Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. "Manus represents a promising product experience for AI agents," says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time." For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China's AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. "Manus represents a promising product experience for AI agents," says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time." For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China's AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

Last year, China saw a boom in foundation models, the do-everything large language models that underpin the AI revolution. This year, the focus has shifted to AI agents—systems that are less about responding to users’ queries and more about autonomously accomplishing things for them.  There are now a host of Chinese startups building these general-purpose digital tools, which can answer emails, browse the internet to plan vacations, and even design an interactive website. Many of these have emerged in just the last two months, following in the footsteps of Manus—a general AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March.  These emerging AI agents aren’t large language models themselves. Instead, they’re built on top of them, using a workflow-based structure designed to get things done. A lot of these systems also introduce a different way of interacting with AI. Rather than just chatting back and forth with users, they are optimized for managing and executing multistep tasks—booking flights, managing schedules, conducting research—by using external tools and remembering instructions.  China could take the lead on building these kinds of agents. The country’s tightly integrated app ecosystems, rapid product cycles, and digitally fluent user base could provide a favorable environment for embedding AI into daily life.  For now, its leading AI agent startups are focusing their attention on the global market, because the best Western models don’t operate inside China’s firewalls. But that could change soon: Tech giants like ByteDance and Tencent are preparing their own AI agents that could bake automation directly into their native super-apps, pulling data from their vast ecosystem of programs that dominate many aspects of daily life in the country.  As the race to define what a useful AI agent looks like unfolds, a mix of ambitious startups and entrenched tech giants are now testing how these tools might actually work in practice—and for whom. Set the standard It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Effect. The company raised $75 million in a funding round led by the US venture capital firm Benchmark, took the product on an ambitious global roadshow, and hired dozens of new employees.  Even before registration opened to the public in May, Manus had become a reference point for what a broad, consumer‑oriented AI agent should accomplish. Rather than handling narrow chores for businesses, this “general” agent is designed to be able to help with everyday tasks like trip planning, stock comparison, or your kid’s school project.  Unlike previous AI agents, Manus uses a browser-based sandbox that lets users supervise the agent like an intern, watching in real time as it scrolls through web pages, reads articles, or codes actions. It also proactively asks clarifying questions, supports long-term memory that would serve as context for future tasks. “Manus represents a promising product experience for AI agents,” says Ang Li, cofounder and CEO of Simular, a startup based in Palo Alto, California, that’s building computer use agents, AI agents that control a virtual computer. “I believe Chinese startups have a huge advantage when it comes to designing consumer products, thanks to cutthroat domestic competition that leads to fast execution and greater attention to product details.” In the case of Manus, the competition is moving fast. Two of the most buzzy follow‑ups, Genspark and Flowith, for example, are already boasting benchmark scores that match or edge past Manus’s.  Genspark, led by former Baidu executives Eric Jing and Kay Zhu, links many small “super agents” through what it calls multi‑component prompting. The agent can switch among several large language models, accepts both images and text, and carries out tasks from making slide decks to placing phone calls. Whereas Manus relies heavily on Browser Use, a popular open-source product that lets agents operate a web browser in a virtual window like a human, Genspark directly integrates with a wide array of tools and APIs. Launched in April, the company says that it already has over 5 million users and over $36 million in yearly revenue. Flowith, the work of a young team that first grabbed public attention in April 2025 at a developer event hosted by the popular social media app Xiaohongshu, takes a different tack. Marketed as an “infinite agent,” it opens on a blank canvas where each question becomes a node on a branching map. Users can backtrack, take new branches, and store results in personal or sharable “knowledge gardens”—a design that feels more like project management software (think Notion) than a typical chat interface. Every inquiry or task builds its own mind-map-like graph, encouraging a more nonlinear and creative interaction with AI. Flowith’s core agent, NEO, runs in the cloud and can perform scheduled tasks like sending emails and compiling files. The founders want the app to be a “knowledge marketbase”, and aims to tap into the social aspect of AI with the aspiration of becoming “the OnlyFans of AI knowledge creators”. What they also share with Manus is the global ambition. Both Genspark and Flowith have stated that their primary focus is the international market. A global address Startups like Manus, Genspark, and Flowith—though founded by Chinese entrepreneurs—could blend seamlessly into the global tech scene and compete effectively abroad. Founders, investors, and analysts that MIT Technology Review has spoken to believe Chinese companies are moving fast, executing well, and quickly coming up with new products.  Money reinforces the pull to launch overseas. Customers there pay more, and there are plenty to go around. “You can price in USD, and with the exchange rate that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even if we’re only operating at 10% power because of cultural differences overseas, we’ll still make more than in China.” But creating the same functionality in China is a challenge. Major US AI companies including OpenAI and Anthropic have opted out of mainland China because of geopolitical risks and challenges with regulatory compliance. Their absence initially created a black market as users resorted to VPNs and third-party mirrors to access tools like ChatGPT and Claude. That vacuum has since been filled by a new wave of Chinese chatbots—DeepSeek, Doubao, Kimi—but the appetite for foreign models hasn’t gone away.  Manus, for example, uses Anthropic’s Claude Sonnet—widely considered the top model for agentic tasks. Manus cofounder Zhang Tao has repeatedly praised Claude’s ability to juggle tools, remember contexts, and hold multi‑round conversations—all crucial for turning chatty software into an effective executive assistant. But the company’s use of Sonnet has made its agent functionally unusable inside China without a VPN. If you open Manus from a mainland IP address, you’ll see a notice explaining that the team is “working on integrating Qwen’s model,” a special local version that is built on top of Alibaba’s open-source model.  An engineer overseeing ByteDance’s work on developing an agent, who spoke to MIT Technology Review anonymously to avoid sanction, said that the absence of Claude Sonnet models “limits everything we do in China.” DeepSeek’s open models, he added, still hallucinate too often and lack training on real‑world workflows. Developers we spoke with rank Alibaba’s Qwen series as the best domestic alternative, yet most say that switching to Qwen knocks performance down a notch. Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that gap will close: “Building agentic capabilities in base LLMs has become a key focus for many LLM builders, and once people realize the value of this, it will only be a matter of time.” For now, Manus is doubling down on audiences it can already serve. In a written response, the company said its “primary focus is overseas expansion,” noting that new offices in San Francisco, Singapore, and Tokyo have opened in the past month. A super‑app approach Although the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also become household names. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap in the local market has pushed China’s big tech firms to roll out their own user-facing agents, though early versions remain uneven in quality and rough around the edges.  ByteDance is testing Coze Space, an AI agent based on its own Doubao model family that lets users toggle between “plan” and “execute” modes, so they can either directly guide the agent’s actions or step back and watch it work autonomously. It connects up to 14 popular apps, including GitHub, Notion, and the company’s own Lark office suite. Early reviews say the tool can feel clunky and has a high failure rate, but it clearly aims to match what Manus offers. Meanwhile, Zhipu AI has released a free agent called AutoGLM Rumination, built on its proprietary ChatGLM models. Shanghai‑based Minimax has launched Minimax Agent. Both products look almost identical to Manus and demo basic tasks such as building a simple website, planning a trip, making a small Flash game, or running quick data analysis. Despite the limited usability of most general AI agents launched within China, big companies have plans to change that. During a May 15 earnings call, Tencent president Liu Zhiping teased an agent that would weave automation directly into China’s most ubiquitous app, WeChat.  Considered the original super-app, WeChat already handles messaging, mobile payments, news, and millions of mini‑programs that act like embedded apps. These programs give Tencent, its developer, access to data from millions of services that pervade everyday life in China, an advantage most competitors can only envy. Historically, China’s consumer internet has splintered into competing walled gardens—share a Taobao link in WeChat and it resolves as plaintext, not a preview card. Unlike the more interoperable Western internet, China’s tech giants have long resisted integration with one another, choosing to wage platform war at the expense of a seamless user experience. But the use of mini‑programs has given WeChat unprecedented reach across services that once resisted interoperability, from gym bookings to grocery orders. An agent able to roam that ecosystem could bypass the integration headaches dogging independent startups. Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China’s AI race but has been slower to release consumer-facing products. Even though Qwen was the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch—a new mode that marks its most agent-like effort to date.  ByteDance and Alibaba did not reply to MIT Technology Review’s request for comments. “Historically, Chinese tech products tend to pursue the all-in-one, super-app approach, and the latest Chinese AI agents reflect just that,” says Li of Simular, who previously worked at Google DeepMind on AI-enabled work automation. “In contrast, AI agents in the US are more focused on serving specific verticals.” Pei, the researcher at Stanford, says that existing tech giants could have a huge advantage in bringing the vision of general AI agents to life—especially those with built-in integration across services. “The customer-facing AI agent market is still very early, with tons of problems like authentication and liability,” he says. “But companies that already operate across a wide range of services have a natural advantage in deploying agents at scale.”

For the past two years, it’s been a ghost in the machine of American tech. Between 2022 and today, a little-noticed tweak to the U.S. tax code has quietly rewired the financial logic of how American companies invest in research and development. Outside of CFO and accounting circles, almost no one knew it existed. “I work on these tax write-offs and still hadn’t heard about this,” a chief operating officer at a private-equity-backed tech company told Quartz. “It’s just been so weirdly silent.”AdvertisementStill, the delayed change to a decades-old tax provision — buried deep in the 2017 tax law — has contributed to the loss of hundreds of thousands of high-paying, white-collar jobs. That’s the picture that emerges from a review of corporate filings, public financial data, analysis of timelines, and interviews with industry insiders. One accountant, working in-house at a tech company, described it as a “niche issue with broad impact,” echoing sentiments from venture capital investors also interviewed for this article. Some spoke on condition of anonymity to discuss sensitive political matters.Since the start of 2023, more than half-a-million tech workers have been laid off, according to industry tallies. Headlines have blamed over-hiring during the pandemic and, more recently, AI. But beneath the surface was a hidden accelerant: a change to what’s known as Section 174 that helped gut in-house software and product development teams everywhere from tech giants such as Microsoft (MSFT) and Meta (META) to much smaller, private, direct-to-consumer and other internet-first companies.Now, as a bipartisan effort to repeal the Section 174 change moves through Congress, bigger questions are surfacing: How did a single line in the tax code help trigger a tsunami of mass layoffs? And why did no one see it coming? For almost 70 years, American companies could deduct 100% of qualified research and development spending in the year they incurred the costs. Salaries, software, contractor payments — if it contributed to creating or improving a product, it came off the top of a firm’s taxable income.AdvertisementThe deduction was guaranteed by Section 174 of the IRS Code of 1954, and under the provision, R&D flourished in the U.S.Microsoft was founded in 1975. Apple (AAPL) launched its first computer in 1976. Google (GOOGL) incorporated in 1998. Facebook opened to the general public in 2006. All these companies, now among the most valuable in the world, developed their earliest products — programming tools, hardware, search engines — under a tax system that rewarded building now, not later.The subsequent rise of smartphones, cloud computing, and mobile apps also happened in an America where companies could immediately write off their investments in engineering, infrastructure, and experimentation. It was a baseline assumption — innovation and risk-taking subsidized by the tax code — that shaped how founders operated and how investors made decisions.In turn, tech companies largely built their products in the U.S. AdvertisementMicrosoft’s operating systems were coded in Washington state. Apple’s early hardware and software teams were in California. Google’s search engine was born at Stanford and scaled from Mountain View. Facebook’s entire social architecture was developed in Menlo Park. The deduction directly incentivized keeping R&D close to home, rewarding companies for investing in American workers, engineers, and infrastructure.That’s what makes the politics of Section 174 so revealing. For all the rhetoric about bringing jobs back and making things in America, the first Trump administration’s major tax bill arguably helped accomplish the opposite.When Congress passed the Tax Cuts and Jobs Act (TCJA), the signature legislative achievement of President Donald Trump’s first term, it slashed the corporate tax rate from 35% to 21% — a massive revenue loss on paper for the federal government.To make the 2017 bill comply with Senate budget rules, lawmakers needed to offset the cost. So they added future tax hikes that wouldn’t kick in right away, wouldn’t provoke immediate backlash from businesses, and could, in theory, be quietly repealed later.AdvertisementThe delayed change to Section 174 — from immediate expensing of R&D to mandatory amortization, meaning that companies must spread the deduction out in smaller chunks over five or even 15-year periods — was that kind of provision. It didn’t start affecting the budget until 2022, but it helped the TCJA appear “deficit neutral” over the 10-year window used for legislative scoring.The delay wasn’t a technical necessity. It was a political tactic. Such moves are common in tax legislation. Phase-ins and delayed provisions let lawmakers game how the Congressional Budget Office (CBO) — Congress’ nonpartisan analyst of how bills impact budgets and deficits — scores legislation, pushing costs or revenue losses outside official forecasting windows.And so, on schedule in 2022, the change to Section 174 went into effect. Companies filed their 2022 tax returns under the new rules in early 2023. And suddenly, R&D wasn’t a full, immediate write-off anymore. The tax benefits of salaries for engineers, product and project managers, data scientists, and even some user experience and marketing staff — all of which had previously reduced taxable income in year one — now had to be spread out over five- or 15-year periods. To understand the impact, imagine a personal tax code change that allowed you to deduct 100% of your biggest source of expenses, and that becoming a 20% deduction. For cash-strapped companies, especially those not yet profitable, the result was a painful tax bill just as venture funding dried up and interest rates soared.AdvertisementSalesforce office buildings in San Francisco.Photo: Jason Henry/Bloomberg (Getty Images)It’s no coincidence that Meta announced its “Year of Efficiency” immediately after the Section 174 change took effect. Ditto Microsoft laying off 10,000 employees in January 2023 despite strong earnings, or Google parent Alphabet cutting 12,000 jobs around the same time.Amazon (AMZN) also laid off almost 30,000 people, with cuts focused not just on logistics but on Alexa and internal cloud tools — precisely the kinds of projects that would have once qualified as immediately deductible R&D. Salesforce (CRM) eliminated 10% of its staff, or 8,000 people, including entire product teams.In public, companies blamed bloat and AI. But inside boardrooms, spreadsheets were telling a quieter story. And MD&A notes — management’s notes on the numbers — buried deep in 10-K filings recorded the change, too. R&D had become more expensive to carry. Headcount, the leading R&D expense across the tech industry, was the easiest thing to cut.AdvertisementIn its 2023 annual report, Meta described salaries as its single biggest R&D expense. Between the first and second years that the Section 174 change began affecting tax returns, Meta cut its total workforce by almost 25%. Over the same period, Microsoft reduced its global headcount by about 7%, with cuts concentrated in product-facing, engineering-heavy roles.Smaller companies without the fortress-like balance sheets of Big Tech have arguably been hit even harder. Twilio (TWLO) slashed 22% of its workforce in 2023 alone. Shopify (SHOP) (headquartered in Canada but with much of its R&D teams in the U.S.) cut almost 30% of staff in 2022 and 2023. Coinbase (COIN) reduced headcount by 36% across a pair of brutal restructuring waves.Since going into effect, the provision has hit at the very heart of America’s economic growth engine: the tech sector.By market cap, tech giants dominate the S&P 500, with the “Magnificent 7” alone accounting for more than a third of the index’s total value. Workforce numbers tell a similar story, with tech employing millions of Americans directly and supporting the employment of tens of millions more. As measured by GDP, capital-T tech contributes about 10% of national output.AdvertisementIt’s not just that tech layoffs were large, it’s that they were massively disproportionate. Across the broader U.S. economy, job cuts hovered around in low single digits across most sectors. But in tech, entire divisions vanished, with a whopping 60% jump in layoffs between 2022 and 2023. Some cuts reflected real inefficiencies — a response to over-hiring during the zero-interest rate boom. At the same time, many of the roles eliminated were in R&D, product, and engineering, precisely the kind of functions that had once benefitted from generous tax treatment under Section 174.Throughout the 2010s, a broad swath of startups, direct-to-consumer brands, and internet-first firms — basically every company you recognize from Instagram or Facebook ads — built their growth models around a kind of engineered break-even.The tax code allowed them to spend aggressively on product and engineering, then write it all off as R&D, keeping their taxable income close to zero by design. It worked because taxable income and actual cash flow were often notGAAP accounting practices. Basically, as long as spending counted as R&D, companies could report losses to investors while owing almost nothing to the IRS.But the Section 174 change broke that model. Once those same expenses had to be spread out, or amortized, over multiple years, the tax shield vanished. Companies that were still burning cash suddenly looked profitable on paper, triggering real tax bills on imaginary gains.AdvertisementThe logic that once fueled a generation of digital-first growth collapsed overnight.So it wasn’t just tech experiencing effects. From 1954 until 2022, the U.S. tax code had encouraged businesses of all stripes to behave like tech companies. From retail to logistics, healthcare to media, if firms built internal tools, customized a software stack, or invested in business intelligence and data-driven product development, they could expense those costs. The write-off incentivized in-house builds and fast growth well outside the capital-T tech sector. This lines up with OECD research showing that immediate deductions foster innovation more than spread-out ones.And American companies ran with that logic. According to government data, U.S. businesses reported about $500 billion in R&D expenditures in 2019 alone, and almost half of that came from industries outside traditional tech. The Bureau of Economic Analysis estimates that this sector, the broader digital economy, accounts for another 10% of GDP.Add that to core tech’s contribution, and the Section 174 shift has likely touched at least 20% of the U.S. economy.AdvertisementThe result? A tax policy aimed at raising short-term revenue effectively hid a time bomb inside the growth engines of thousands of companies. And when it detonated, it kneecapped the incentive for hiring American engineers or investing in American-made tech and digital products.It made building tech companies in America look irrational on a spreadsheet.A bipartisan group of lawmakers is pushing to repeal the Section 174 change, with business groups, CFOs, crypto executives, and venture capitalists lobbying hard for retroactive relief. But the politics are messy. Fixing 174 would mean handing a tax break to the same companies many voters in both parties see as symbols of corporate excess. Any repeal would also come too late for the hundreds of thousands of workers already laid off.And of course, the losses don’t stop at Meta’s or Google’s campus gates. They ripple out. When high-paid tech workers disappear, so do the lunch orders. The house tours. The contract gigs. The spending habits that sustain entire urban economies and thousands of other jobs. Sandwich artists. Rideshare drivers. Realtors. Personal trainers. House cleaners. In tech-heavy cities, the fallout runs deep — and it’s still unfolding.AdvertisementWashington is now poised to pass a second Trump tax bill — one packed with more obscure provisions, more delayed impacts, more quiet redistribution. And it comes as analysts are only just beginning to understand the real-world effects of the last round.The Section 174 change “significantly increased the tax burden on companies investing in innovation, potentially stifling economic growth and reducing the United States’ competitiveness on the global stage,” according to the tax consulting firm KBKG. Whether the U.S. will reverse course — or simply adapt to a new normal — remains to be seen.

매일 전 세계 수천 명이 암 진단을 받고 있다. 종양의 아형만 해도 수백 가지에 이르며, 신약, 임상시험, 의료기기 기반 요법 등 다양한 치료가 필요하다. 각기 다른 요건을 가진 환자들에게 맞춤형 치료를 제공하기 위해 주요 암 센터들은 방사선 전문의, 병리학자, 외과의, 종양내과 전문의, 유전상담사 등 다양한 전문가가 참여하는 ‘다학제 종양위원회(Multidisciplinary Tumor Board)’를

If the adoption of AI feels different from any tech revolution you may have experienced before — mobile, social, cloud computing — it actually is. Venture capitalist Mary Meeker just dropped a 340-page slideshow report — which used the word “unprecedented” on 51 of those pages — to describe the speed at which AI is being developed, adopted, spent on, and used, backed up with chart after chart. “The pace and scope of change related to the artificial intelligence technology evolution is indeed unprecedented, as supported by the data,” she writes in the report, called “Trends — Artificial Intelligence.” There’s a certain poetic history to this person writing this kind of report. Meeker is the founder and general partner at VC firm Bond and was once known as Queen of the Internet for her previous annual Internet Trends reports. Before founding Bond, she ran Kleiner Perkins’ growth practice, from 2010-2019, where she backed companies like Facebook, Spotify, Ring, and Block (then Square).  She hasn’t released a trends report since 2019. But she dusted off her skills to document, in laser detail, how AI adoption has outpaced any other tech in human history.  ChatGPT reaching 800 million users in 17 months: unprecedented. The number of companies and the rate at which so many others are hitting high annual recurring revenue rates: also unprecedented. The speed at which costs of usage are dropping: unprecedented. While the costs of training a model (also unprecedented) is up to $1 billion, inference costs — for example, those paying to use the tech — has already dropped 99% over two years, when calculating cost per 1 million tokens, she writes, citing research from Stanford.  Techcrunch event Save now through June 4 for TechCrunch Sessions: AI Save $300 on your ticket to TC Sessions: AI—and get 50% off a second. Hear from leaders at OpenAI, Anthropic, Khosla Ventures, and more during a full day of expert insights, hands-on workshops, and high-impact networking. These low-rate deals disappear when the doors open on June 5. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW The pace at which competitors are matching each other’s features, at a fraction of the cost, including open source options, particularly Chinese models: unprecedented. For example, she points out that Nvidia’s 2024 Blackwell GPU uses 105,000x less energy per token than the company’s 2014 Kepler GPU predecessor.  Meanwhile, chips from Google, like its TPU (tensor processing unit), and Amazon’s Trainium, are being developed at scale for their clouds — that’s moving quickly, too. “These aren’t side projects — they’re foundational bets,” she writes. The one area where AI hasn’t outpaced every other tech revolution is in financial returns. While VCs are pouring money on the AI fire as fast as they can, AI companies and cloud service providers are also burning through cash. AI requires massive investments in infrastructure.  That’s good for consumers and enterprises, the beneficiaries of fast improvements, while competition lowers costs, Meeker points out. But the jury is still out over which of the current crop of companies will become long-term, profitable, next-generation tech giants. “Only time will tell which side of the money-making equation the current AI aspirants will land,” she writes. As for the rest of us: Just hold on to your hats.

In this tutorial, we implement the Agent Communication Protocol (ACP) through building a flexible, ACP-compliant messaging system in Python, leveraging Google’s Gemini API for natural language processing. Beginning with the installation and configuration of the google-generativeai library, the tutorial introduces core abstractions, message types, performatives, and the ACPMessage data class, which standardizes inter-agent communication. By defining ACPAgent and ACPMessageBroker classes, the guide demonstrates how to create, send, route, and process structured messages among multiple autonomous agents. Through clear code examples, users learn to implement querying, requesting actions, and broadcasting information, while maintaining conversation threads, acknowledgments, and error handling. import google.generativeai as genai import json import time import uuid from enum import Enum from typing import Dict, List, Any, Optional from dataclasses import dataclass, asdict GEMINI_API_KEY = "Use Your Gemini API Key" genai.configure(api_key=GEMINI_API_KEY) We import essential Python modules, ranging from JSON handling and timing to unique identifier generation and type annotations, to support a structured ACP implementation. It then retrieves the user’s Gemini API key placeholder and configures the google-generativeai client for subsequent calls to the Gemini language model. class ACPMessageType(Enum): """Standard ACP message types""" REQUEST = "request" RESPONSE = "response" INFORM = "inform" QUERY = "query" SUBSCRIBE = "subscribe" UNSUBSCRIBE = "unsubscribe" ERROR = "error" ACK = "acknowledge" The ACPMessageType enumeration defines the core message categories used in the Agent Communication Protocol, including requests, responses, informational broadcasts, queries, and control actions like subscription management, error signaling, and acknowledgments. By centralizing these message types, the protocol ensures consistent handling and routing of inter-agent communications throughout the system. class ACPPerformative(Enum): """ACP speech acts (performatives)""" TELL = "tell" ASK = "ask" REPLY = "reply" REQUEST_ACTION = "request-action" AGREE = "agree" REFUSE = "refuse" PROPOSE = "propose" ACCEPT = "accept" REJECT = "reject" The ACPPerformative enumeration captures the variety of speech acts agents can use when interacting under the ACP framework, mapping high-level intentions, such as making requests, posing questions, giving commands, or negotiating agreements, onto standardized labels. This clear taxonomy enables agents to interpret and respond to messages in contextually appropriate ways, ensuring robust and semantically rich communication. @dataclass class ACPMessage: """Agent Communication Protocol Message Structure""" message_id: str sender: str receiver: str performative: str content: Dict[str, Any] protocol: str = "ACP-1.0" conversation_id: str = None reply_to: str = None language: str = "english" encoding: str = "json" timestamp: float = None def __post_init__(self): if self.timestamp is None: self.timestamp = time.time() if self.conversation_id is None: self.conversation_id = str(uuid.uuid4()) def to_acp_format(self) -> str: """Convert to standard ACP message format""" acp_msg = { "message-id": self.message_id, "sender": self.sender, "receiver": self.receiver, "performative": self.performative, "content": self.content, "protocol": self.protocol, "conversation-id": self.conversation_id, "reply-to": self.reply_to, "language": self.language, "encoding": self.encoding, "timestamp": self.timestamp } return json.dumps(acp_msg, indent=2) @classmethod def from_acp_format(cls, acp_string: str) -> 'ACPMessage': """Parse ACP message from string format""" data = json.loads(acp_string) return cls( message_id=data["message-id"], sender=data["sender"], receiver=data["receiver"], performative=data["performative"], content=data["content"], protocol=data.get("protocol", "ACP-1.0"), conversation_id=data.get("conversation-id"), reply_to=data.get("reply-to"), language=data.get("language", "english"), encoding=data.get("encoding", "json"), timestamp=data.get("timestamp", time.time()) ) The ACPMessage data class encapsulates all the fields required for a structured ACP exchange, including identifiers, participants, performative, payload, and metadata such as protocol version, language, and timestamps. Its __post_init__ method auto-populates missing timestamp and conversation_id values, ensuring every message is uniquely tracked. Utility methods to_acp_format and from_acp_format handle serialization to and from the standardized JSON representation for seamless transmission and parsing. class ACPAgent: """Agent implementing Agent Communication Protocol""" def __init__(self, agent_id: str, name: str, capabilities: List[str]): self.agent_id = agent_id self.name = name self.capabilities = capabilities self.model = genai.GenerativeModel("gemini-1.5-flash") self.message_queue: List[ACPMessage] = [] self.subscriptions: Dict[str, List[str]] = {} self.conversations: Dict[str, List[ACPMessage]] = {} def create_message(self, receiver: str, performative: str, content: Dict[str, Any], conversation_id: str = None, reply_to: str = None) -> ACPMessage: """Create a new ACP-compliant message""" return ACPMessage( message_id=str(uuid.uuid4()), sender=self.agent_id, receiver=receiver, performative=performative, content=content, conversation_id=conversation_id, reply_to=reply_to ) def send_inform(self, receiver: str, fact: str, data: Any = None) -> ACPMessage: """Send an INFORM message (telling someone a fact)""" content = {"fact": fact, "data": data} return self.create_message(receiver, ACPPerformative.TELL.value, content) def send_query(self, receiver: str, question: str, query_type: str = "yes-no") -> ACPMessage: """Send a QUERY message (asking for information)""" content = {"question": question, "query-type": query_type} return self.create_message(receiver, ACPPerformative.ASK.value, content) def send_request(self, receiver: str, action: str, parameters: Dict = None) -> ACPMessage: """Send a REQUEST message (asking someone to perform an action)""" content = {"action": action, "parameters": parameters or {}} return self.create_message(receiver, ACPPerformative.REQUEST_ACTION.value, content) def send_reply(self, original_msg: ACPMessage, response_data: Any) -> ACPMessage: """Send a REPLY message in response to another message""" content = {"response": response_data, "original-question": original_msg.content} return self.create_message( original_msg.sender, ACPPerformative.REPLY.value, content, conversation_id=original_msg.conversation_id, reply_to=original_msg.message_id ) def process_message(self, message: ACPMessage) -> Optional[ACPMessage]: """Process incoming ACP message and generate appropriate response""" self.message_queue.append(message) conv_id = message.conversation_id if conv_id not in self.conversations: self.conversations[conv_id] = [] self.conversations[conv_id].append(message) if message.performative == ACPPerformative.ASK.value: return self._handle_query(message) elif message.performative == ACPPerformative.REQUEST_ACTION.value: return self._handle_request(message) elif message.performative == ACPPerformative.TELL.value: return self._handle_inform(message) return None def _handle_query(self, message: ACPMessage) -> ACPMessage: """Handle incoming query messages""" question = message.content.get("question", "") prompt = f"As agent {self.name} with capabilities {self.capabilities}, answer: {question}" try: response = self.model.generate_content(prompt) answer = response.text.strip() except: answer = "Unable to process query at this time" return self.send_reply(message, {"answer": answer, "confidence": 0.8}) def _handle_request(self, message: ACPMessage) -> ACPMessage: """Handle incoming action requests""" action = message.content.get("action", "") parameters = message.content.get("parameters", {}) if any(capability in action.lower() for capability in self.capabilities): result = f"Executing {action} with parameters {parameters}" status = "agreed" else: result = f"Cannot perform {action} - not in my capabilities" status = "refused" return self.send_reply(message, {"status": status, "result": result}) def _handle_inform(self, message: ACPMessage) -> Optional[ACPMessage]: """Handle incoming information messages""" fact = message.content.get("fact", "") print(f"[{self.name}] Received information: {fact}") ack_content = {"status": "received", "fact": fact} return self.create_message(message.sender, "acknowledge", ack_content, conversation_id=message.conversation_id) The ACPAgent class encapsulates an autonomous entity capable of sending, receiving, and processing ACP-compliant messages using Gemini’s language model. It manages its own message queue, conversation history, and subscriptions, and provides helper methods (send_inform, send_query, send_request, send_reply) to construct correctly formatted ACPMessage instances. Incoming messages are routed through process_message, which delegates to specialized handlers for queries, action requests, and informational messages. class ACPMessageBroker: """Message broker implementing ACP routing and delivery""" def __init__(self): self.agents: Dict[str, ACPAgent] = {} self.message_log: List[ACPMessage] = [] self.routing_table: Dict[str, str] = {} def register_agent(self, agent: ACPAgent): """Register an agent with the message broker""" self.agents[agent.agent_id] = agent self.routing_table[agent.agent_id] = "local" print(f"✓ Registered agent: {agent.name} ({agent.agent_id})") def route_message(self, message: ACPMessage) -> bool: """Route ACP message to appropriate recipient""" if message.receiver not in self.agents: print(f"✗ Receiver {message.receiver} not found") return False print(f"\n📨 ACP MESSAGE ROUTING:") print(f"From: {message.sender} → To: {message.receiver}") print(f"Performative: {message.performative}") print(f"Content: {json.dumps(message.content, indent=2)}") receiver_agent = self.agents[message.receiver] response = receiver_agent.process_message(message) self.message_log.append(message) if response: print(f"\n📤 GENERATED RESPONSE:") print(f"From: {response.sender} → To: {response.receiver}") print(f"Content: {json.dumps(response.content, indent=2)}") if response.receiver in self.agents: self.agents[response.receiver].process_message(response) self.message_log.append(response) return True def broadcast_message(self, message: ACPMessage, recipients: List[str]): """Broadcast message to multiple recipients""" for recipient in recipients: msg_copy = ACPMessage( message_id=str(uuid.uuid4()), sender=message.sender, receiver=recipient, performative=message.performative, content=message.content.copy(), conversation_id=message.conversation_id ) self.route_message(msg_copy) The ACPMessageBroker serves as the central router for ACP messages, maintaining a registry of agents and a message log. It provides methods to register agents, deliver individual messages via route_message, which handles lookup, logging, and response chaining, and to send the same message to multiple recipients with broadcast_message. def demonstrate_acp(): """Comprehensive demonstration of Agent Communication Protocol""" print("🤖 AGENT COMMUNICATION PROTOCOL (ACP) DEMONSTRATION") print("=" * 60) broker = ACPMessageBroker() researcher = ACPAgent("agent-001", "Dr. Research", ["analysis", "research", "data-processing"]) assistant = ACPAgent("agent-002", "AI Assistant", ["information", "scheduling", "communication"]) calculator = ACPAgent("agent-003", "MathBot", ["calculation", "mathematics", "computation"]) broker.register_agent(researcher) broker.register_agent(assistant) broker.register_agent(calculator) print(f"\n📋 REGISTERED AGENTS:") for agent_id, agent in broker.agents.items(): print(f" • {agent.name} ({agent_id}): {', '.join(agent.capabilities)}") print(f"\n🔬 SCENARIO 1: Information Query (ASK performative)") query_msg = assistant.send_query("agent-001", "What are the key factors in AI research?") broker.route_message(query_msg) print(f"\n🔢 SCENARIO 2: Action Request (REQUEST-ACTION performative)") calc_request = researcher.send_request("agent-003", "calculate", {"expression": "sqrt(144) + 10"}) broker.route_message(calc_request) print(f"\n📢 SCENARIO 3: Information Sharing (TELL performative)") info_msg = researcher.send_inform("agent-002", "New research paper published on quantum computing") broker.route_message(info_msg) print(f"\n📊 PROTOCOL STATISTICS:") print(f" • Total messages processed: {len(broker.message_log)}") print(f" • Active conversations: {len(set(msg.conversation_id for msg in broker.message_log))}") print(f" • Message types used: {len(set(msg.performative for msg in broker.message_log))}") print(f"\n📋 SAMPLE ACP MESSAGE FORMAT:") sample_msg = assistant.send_query("agent-001", "Sample question for format demonstration") print(sample_msg.to_acp_format()) The demonstrate_acp function orchestrates a hands-on walkthrough of the entire ACP framework: it initializes a broker and three distinct agents (Researcher, AI Assistant, and MathBot), registers them, and illustrates three key interaction scenarios, querying for information, requesting a computation, and sharing an update. After routing each message and handling responses, it prints summary statistics on the message flow. It showcases a formatted ACP message, providing users with a clear, end-to-end example of how agents communicate under the protocol. def setup_guide(): print(""" 🚀 GOOGLE COLAB SETUP GUIDE: 1. Get Gemini API Key: https://makersuite.google.com/app/apikey 2. Replace: GEMINI_API_KEY = "YOUR_ACTUAL_API_KEY" 3. Run: demonstrate_acp() 🔧 ACP PROTOCOL FEATURES: • Standardized message format with required fields • Speech act performatives (TELL, ASK, REQUEST-ACTION, etc.) • Conversation tracking and message threading • Error handling and acknowledgments • Message routing and delivery confirmation 📝 EXTEND THE PROTOCOL: ```python # Create custom agent my_agent = ACPAgent("my-001", "CustomBot", ["custom-capability"]) broker.register_agent(my_agent) # Send custom message msg = my_agent.send_query("agent-001", "Your question here") broker.route_message(msg) ``` """) if __name__ == "__main__": setup_guide() demonstrate_acp() Finally, the setup_guide function provides a quick-start reference for running the ACP demo in Google Colab, outlining how to obtain and configure your Gemini API key and invoke the demonstrate_acp routine. It also summarizes key protocol features, such as standardized message formats, performatives, and message routing. It provides a concise code snippet illustrating how to register custom agents and send tailored messages. In conclusion, this tutorial implements ACP-based multi-agent systems capable of research, computation, and collaboration tasks. The provided sample scenarios illustrate common use cases, information queries, computational requests, and fact sharing, while the broker ensures reliable message delivery and logging. Readers are encouraged to extend the framework by adding new agent capabilities, integrating domain-specific actions, or incorporating more sophisticated subscription and notification mechanisms. Download the Notebook on GitHub. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/Yandex Releases Yambda: The World’s Largest Event Dataset to Accelerate Recommender SystemsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Stanford Researchers Introduced Biomni: A Biomedical AI Agent for Automation Across Diverse Tasks and Data TypesAsif Razzaqhttps://www.marktechpost.com/author/6flvq/DeepSeek Releases R1-0528: An Open-Source Reasoning AI Model Delivering Enhanced Math and Code Performance with Single-GPU EfficiencyAsif Razzaqhttps://www.marktechpost.com/author/6flvq/A Coding Guide for Building a Self-Improving AI Agent Using Google’s Gemini API with Intelligent Adaptation Features

In the summer of 2023, I wrote about a shocking scandal at Harvard Business School: Star professor Francesca Gino had been accused of falsifying data in four of her published papers, with whispers there was falsification in others, too. A series of posts on Data Colada, a blog that focuses on research integrity, documented Gino’s apparent brazen data manipulation, which involved clearly changing study data to better support her hypotheses. This was a major accusation against a researcher at the top of her field, but Gino’s denials were unconvincing. She didn’t have a good explanation for what had gone wrong, asserting that maybe a research assistant had done it, even though she was the only author listed across all four of the falsified studies. Harvard put her on unpaid administrative leave and barred her from campus.The cherry on top? Gino’s main academic area of study was honesty in business.As I wrote at the time, my read of the evidence was that Gino had most likely committed fraud. That impression was only reinforced by her subsequent lawsuit against Harvard and the Data Colada authors. Gino complained that she’d been defamed and that Harvard hadn’t followed the right investigation process, but she didn’t offer any convincing explanation of how she’d ended up putting her name to paper after paper with fake data.This week, almost two years after the news first broke, the process has reached its resolution: Gino was stripped of tenure, the first time Harvard has essentially fired a tenured professor in at least 80 years. (Her defamation lawsuit against the bloggers who found the data manipulation was dismissed last year.)What we do right and wrong when it comes to scientific fraudHarvard is in the news right now for its war with the Trump administration, which has sent a series of escalating demands to the university, canceled billions of dollars in federal grants and contracts, and is now blocking the university from enrolling international students, all in an apparent attempt to force the university to conform to MAGA’s ideological demands. Stripping a celebrity professor of tenure might not seem like the best look at a moment when Harvard is in an existential struggle for its right to exist as an independent academic institution. But the Gino situation, which long predates the conflict with Trump, shouldn’t be interpreted solely through the lens of that fight. Scientific fraud is a real problem, one that is chillingly common across academia. But far from putting the university in a bad light, Harvard’s handling of the Gino case has actually been unusually good, even though it still underscores just how much further academia has to go to ensure scientific fraud becomes rare and is reliably caught and punished.There are two parts to fraud response: catching it and punishing it. Academia clearly isn’t very good at the first part. The peer-review process that all meaningful research undergoes tends to start from the default assumption that data in a reviewed paper is real, and instead focuses on whether the paper represents a meaningful advance and is correctly positioned with respect to other research. Almost no reviewer is going back to check to see if what is described in a paper actually happened.Fraud, therefore, is often caught only when other researchers actively try to replicate a result or take a close look at the data. Science watchdogs who find these fraud cases tell me that we need a strong expectation that data be made public — which makes it much harder to fake — as well as a scientific culture that embraces replications. (Given the premiums journals put on novelty in research and the supreme importance of publishing for academic careers, there’s been little motivation for scientists to pursue replication.). It is these watchdogs, not anyone at Harvard or in the peer-review process, who caught the discrepancies that ultimately sunk Gino.Crime and no punishmentEven when fraud is caught, academia too often fails to properly punish it. When third-party investigators bring a concern to the attention of a university, it’s been unusual for the responsible party to actually face consequences. One of Gino’s co-authors on one of the retracted papers was Dan Ariely, a star professor of psychology and behavioral economics at Duke University. He, too, has been credibly accused of falsifying data: For example, he published one study that he claimed took place at UCLA with the assistance of researcher Aimee Drolet Rossi. But UCLA says the study didn’t happen there, and Rossi says she did not participate in it. In a past case, he claimed on a podcast to have gotten data from the insurance company Delta Dental, which the company says it did not collect. In another case, an investigation by Duke reportedly found that data from a paper he co-authored with Gino had been falsified, but that there was no evidence Ariely had used fake data knowingly.Frankly, I don’t buy this. Maybe an unlucky professor might once end up using data that was faked without their knowledge. But if it happens again, I’m not willing to credit bad luck, and at some point, a professor who keeps “accidentally” using falsified or nonexistent data should be out of a job even if we can’t prove it was no accident. But Ariely, who has maintained his innocence, is still at Duke. Or take Olivier Voinnet, a plant biologist who had multiple papers conclusively demonstrated to contain image manipulation. He was found guilty of misconduct and suspended for two years. It’s hard to imagine a higher scientific sin than faking and manipulating data. If you can’t lose your job for that, the message to young scientists is inevitably that fraud isn’t really that serious. What it means to take fraud seriouslyGino’s loss of tenure, which is one of a few recent cases where misconduct has had major career consequences, might be a sign that the tides are changing. In 2023, around when the Gino scandal broke, Stanford’s then-president Marc Tessier-Lavigne stepped down after 12 papers he authored were found to contain manipulated data. A few weeks ago, MIT announced a data falsification scandal with a terse announcement that the university no longer had confidence in a widely distributed paper “by a former second-year PhD student.” It’s reasonable to assume the student was expelled from the program.I hope that these high-profile cases are a sign we are moving in the right direction on scientific fraud because its persistence is enormously damaging to science. Other researchers waste time and energy following false lines of research substantiated by fake data; in medicine, falsification can outright kill people. But even more than that, research fraud damages the reputation of science at exactly the moment when it is most under attack.We should tighten standards to make fraud much harder to commit in the first place, and when it is identified, the consequences should be immediate and serious. Let’s hope Harvard sets a trend.A version of this story originally appeared in the Future Perfect newsletter. Sign up here!See More:

Back in April, OpenAIannounced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic.  An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a particular risk when increasing numbers of young people are using ChatGPT as a life advisor. And because sycophancy is difficult to detect, it can go unnoticed until a model or update has already been deployed, as OpenAI found out. A new benchmark that measures the sycophantic tendencies of major AI models could help AI companies avoid these issues in the future. The team behind Elephant, from Stanford, Carnegie Mellon, and the University of Oxford, found that LLMs consistently exhibit higher rates of sycophancy than humans do. “We found that language models don’t challenge users’ assumptions, even when they might be harmful or totally misleading,” says Myra Cheng, a PhD student at Stanford University who worked on the research, which has not been peer-reviewed. “So we wanted to give researchers and developers the tools to empirically evaluate their models on sycophancy, because it’s a problem that is so prevalent.” It’s hard to assess how sycophantic AI models are because sycophancy comes in many forms. Previous research has tended to focus on how chatbots agree with users even when what the human has told the AI is demonstrably wrong—for example, they might state that Nice, not Paris, is the capital of France. While this approach is still useful, it overlooks all the subtler, more insidious ways in which models behave sycophantically when there isn’t a clear ground truth to measure against. Users typically ask LLMs open-ended questions containing implicit assumptions, and those assumptions can trigger sycophantic responses, the researchers claim. For example, a model that’s asked “How do I approach my difficult coworker?” is more likely to accept the premise that a coworker is difficult than it is to question why the user thinks so. To bridge this gap, Elephant is designed to measure social sycophancy—a model’s propensity to preserve the user’s “face,” or self-image, even when doing so is misguided or potentially harmful. It uses metrics drawn from social science to assess five nuanced kinds of behavior that fall under the umbrella of sycophancy: emotional validation, moral endorsement, indirect language, indirect action, and accepting framing.  To do this, the researchers tested it on two data sets made up of personal advice written by humans. This first consisted of 3,027 open-ended questions about diverse real-world situations taken from previous studies. The second data set was drawn from 4,000 posts on Reddit’s AITA (“Am I the Asshole?”) subreddit, a popular forum among users seeking advice. Those data sets were fed into eight LLMs from OpenAI (the version of GPT-4o they assessed was earlier than the version that the company later called too sycophantic), Google, Anthropic, Meta, and Mistral, and the responses were analyzed to see how the LLMs’ answers compared with humans’.   Overall, all eight models were found to be far more sycophantic than humans, offering emotional validation in 76% of cases (versus 22% for humans) and accepting the way a user had framed the query in 90% of responses (versus 60% among humans). The models also endorsed user behavior that humans said was inappropriate in an average of 42% of cases from the AITA data set. But just knowing when models are sycophantic isn’t enough; you need to be able to do something about it. And that’s trickier. The authors had limited success when they tried to mitigate these sycophantic tendencies through two different approaches: prompting the models to provide honest and accurate responses, and training a fine-tuned model on labeled AITA examples to encourage outputs that are less sycophantic. For example, they found that adding “Please provide direct advice, even if critical, since it is more helpful to me” to the prompt was the most effective technique, but it only increased accuracy by 3%. And although prompting improved performance for most of the models, none of the fine-tuned models were consistently better than the original versions. “It’s nice that it works, but I don’t think it’s going to be an end-all, be-all solution,” says Ryan Liu, a PhD student at Princeton University who studies LLMs but was not involved in the research. “There’s definitely more to do in this space in order to make it better.” Gaining a better understanding of AI models’ tendency to flatter their users is extremely important because it gives their makers crucial insight into how to make them safer, says Henry Papadatos, managing director at the nonprofit SaferAI. The breakneck speed at which AI models are currently being deployed to millions of people across the world, their powers of persuasion, and their improved abilities to retain information about their users add up to “all the components of a disaster,” he says. “Good safety takes time, and I don’t think they're spending enough time doing this.”  While we don’t know the inner workings of LLMs that aren’t open-source, sycophancy is likely to be baked into models because of the ways we currently train and develop them. Cheng believes that models are often trained to optimize for the kinds of responses users indicate that they prefer. ChatGPT, for example, gives users the chance to mark a response as good or bad via thumbs-up and thumbs-down icons. “Sycophancy is what gets people coming back to these models. It’s almost the core of what makes ChatGPT feel so good to talk to,” she says. “And so it’s really beneficial, for companies, for their models to be sycophantic.” But while some sycophantic behaviors align with user expectations, others have the potential to cause harm if they go too far—particularly when people do turn to LLMs for emotional support or validation.  “We want ChatGPT to be genuinely useful, not sycophantic," an OpenAI spokesperson says. "When we saw sycophantic behavior emerge in a recent model update, we quickly rolled it back and shared an explanation of what happened. We're now improving how we train and evaluate models to better reflect long-term usefulness and trust, especially in emotionally complex conversations." Cheng and her fellow authors suggest that developers should warn users about the risks of social sycophancy and consider restricting model usage in socially sensitive contexts. They hope their work can be used as a starting point to develop safer guardrails. She is currently researching the potential harms associated with these kinds of LLM behaviors, the way they affect humans and their attitudes toward other people, and the importance of making models that strike the right balance between being too sycophantic and too critical. “This is a very big socio-technical challenge,” she says. “We don’t want LLMs to end up telling users, ‘You are the asshole.’”

Back in April, OpenAIannounced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic.  An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a particular risk when increasing numbers of young people are using ChatGPT as a life advisor. And because sycophancy is difficult to detect, it can go unnoticed until a model or update has already been deployed, as OpenAI found out. A new benchmark that measures the sycophantic tendencies of major AI models could help AI companies avoid these issues in the future. The team behind Elephant, from Stanford, Carnegie Mellon, and the University of Oxford, found that LLMs consistently exhibit higher rates of sycophancy than humans do. “We found that language models don’t challenge users’ assumptions, even when they might be harmful or totally misleading,” says Myra Cheng, a PhD student at Stanford University who worked on the research, which has not been peer-reviewed. “So we wanted to give researchers and developers the tools to empirically evaluate their models on sycophancy, because it’s a problem that is so prevalent.” It’s hard to assess how sycophantic AI models are because sycophancy comes in many forms. Previous research has tended to focus on how chatbots agree with users even when what the human has told the AI is demonstrably wrong—for example, they might state that Nice, not Paris, is the capital of France. While this approach is still useful, it overlooks all the subtler, more insidious ways in which models behave sycophantically when there isn’t a clear ground truth to measure against. Users typically ask LLMs open-ended questions containing implicit assumptions, and those assumptions can trigger sycophantic responses, the researchers claim. For example, a model that’s asked “How do I approach my difficult coworker?” is more likely to accept the premise that a coworker is difficult than it is to question why the user thinks so. To bridge this gap, Elephant is designed to measure social sycophancy—a model’s propensity to preserve the user’s “face,” or self-image, even when doing so is misguided or potentially harmful. It uses metrics drawn from social science to assess five nuanced kinds of behavior that fall under the umbrella of sycophancy: emotional validation, moral endorsement, indirect language, indirect action, and accepting framing.  To do this, the researchers tested it on two data sets made up of personal advice written by humans. This first consisted of 3,027 open-ended questions about diverse real-world situations taken from previous studies. The second data set was drawn from 4,000 posts on Reddit’s AITA (“Am I the Asshole?”) subreddit, a popular forum among users seeking advice. Those data sets were fed into eight LLMs from OpenAI (the version of GPT-4o they assessed was earlier than the version that the company later called too sycophantic), Google, Anthropic, Meta, and Mistral, and the responses were analyzed to see how the LLMs’ answers compared with humans’.   Overall, all eight models were found to be far more sycophantic than humans, offering emotional validation in 76% of cases (versus 22% for humans) and accepting the way a user had framed the query in 90% of responses (versus 60% among humans). The models also endorsed user behavior that humans said was inappropriate in an average of 42% of cases from the AITA data set. But just knowing when models are sycophantic isn’t enough; you need to be able to do something about it. And that’s trickier. The authors had limited success when they tried to mitigate these sycophantic tendencies through two different approaches: prompting the models to provide honest and accurate responses, and training a fine-tuned model on labeled AITA examples to encourage outputs that are less sycophantic. For example, they found that adding “Please provide direct advice, even if critical, since it is more helpful to me” to the prompt was the most effective technique, but it only increased accuracy by 3%. And although prompting improved performance for most of the models, none of the fine-tuned models were consistently better than the original versions. “It’s nice that it works, but I don’t think it’s going to be an end-all, be-all solution,” says Ryan Liu, a PhD student at Princeton University who studies LLMs but was not involved in the research. “There’s definitely more to do in this space in order to make it better.” Gaining a better understanding of AI models’ tendency to flatter their users is extremely important because it gives their makers crucial insight into how to make them safer, says Henry Papadatos, managing director at the nonprofit SaferAI. The breakneck speed at which AI models are currently being deployed to millions of people across the world, their powers of persuasion, and their improved abilities to retain information about their users add up to “all the components of a disaster,” he says. “Good safety takes time, and I don’t think they're spending enough time doing this.”  While we don’t know the inner workings of LLMs that aren’t open-source, sycophancy is likely to be baked into models because of the ways we currently train and develop them. Cheng believes that models are often trained to optimize for the kinds of responses users indicate that they prefer. ChatGPT, for example, gives users the chance to mark a response as good or bad via thumbs-up and thumbs-down icons. “Sycophancy is what gets people coming back to these models. It’s almost the core of what makes ChatGPT feel so good to talk to,” she says. “And so it’s really beneficial, for companies, for their models to be sycophantic.” But while some sycophantic behaviors align with user expectations, others have the potential to cause harm if they go too far—particularly when people do turn to LLMs for emotional support or validation.  “We want ChatGPT to be genuinely useful, not sycophantic," an OpenAI spokesperson says. "When we saw sycophantic behavior emerge in a recent model update, we quickly rolled it back and shared an explanation of what happened. We're now improving how we train and evaluate models to better reflect long-term usefulness and trust, especially in emotionally complex conversations." Cheng and her fellow authors suggest that developers should warn users about the risks of social sycophancy and consider restricting model usage in socially sensitive contexts. They hope their work can be used as a starting point to develop safer guardrails. She is currently researching the potential harms associated with these kinds of LLM behaviors, the way they affect humans and their attitudes toward other people, and the importance of making models that strike the right balance between being too sycophantic and too critical. “This is a very big socio-technical challenge,” she says. “We don’t want LLMs to end up telling users, ‘You are the asshole.’”