AI Agents

Google said on Nov. 11 (Nov. 12 Japan time) on its official X account that developers can now connect sandboxes (isolated execution environments) they set up themselves to Managed Agents, the AI agent feature of the Gemini API. Tools that had run in Linux sandboxes provided by Google can now run on a company's own computing infrastructure. The connection uses outbound HTTPS from the open-source Gemini API CLI, so no inbound firewall ports are needed. The feature is in preview, and Google says its specifications may change. The official documentation does not mention pricing for computing resources in self-hosted environments.

SpaceXSI (SpaceX Super Intelligence) has released version "1.0" of its AI agent "Grok Bot." Its official news post calls it "a major update to our AI teammates." The update adds "Team Bots," which let a team share the same bot; collaboration in Slack and Microsoft Teams; joining meetings on Google Meet, Zoom and Teams; proactive offers to take on work; and learning a team's preferences and decisions. Team Bots are available on the Team and Enterprise plans, and meetings on the Team, Enterprise and Ultra plans.

China's Alibaba on Nov. 11 released "1.0 Preview," the next version of its AI programming agent "Qwen Code," as a test release on GitHub. It merged the agent foundations of its commercial AI coding service "Qoder" and Qwen Code into one and released it as open source under the Apache 2.0 license. Besides Qwen and Qoder accounts, users can also use other companies' models and their own endpoints. Packages for macOS, Linux and Windows run without installing Node.js separately. The company said this is a pre-release version and that the interface, settings and behavior may change. Users have also reported dropped connections.

AI coding tool maker Cursor on Nov. 10 announced in beta that it has rebuilt "Projects," which coordinates multiple AI agents to carry out large development work, so team members can use it together. It has a shared task board and shared materials such as requirements documents, and each member's agents coordinate overlaps and ordering with one another's work. Sharing requires a team account and Cursor 3.26.13 or later, and it is not available on the Free and Plus plans. The company has not disclosed pricing, limits on the number of people or how accurate the coordination is.

OpenAI said on Nov. 6 (Nov. 7 Japan time) on its official X account for developers that it had added a "Showcase" tab to the "OpenAI Developers" plugin used in ChatGPT and Codex. Eligible ChatGPT Pro and Business subscribers can apply API credits claimed in billing settings to their app's API usage, it said. According to official documentation, Pro credits match the subscription price at $100 to $500 a month, while Business credits add up per-seat amounts to a maximum of $500 a month, and unused credits expire at the end of the billing period. Codex usage via API keys is not covered. The official documentation does not describe the tab's interface or contents, and they could not be confirmed.

U.S.-based AI model routing service OpenRouter on Nov. 6 (Nov. 7 Japan time) rebuilt "Auto Router," which automatically chooses the model to use for each request, and launched "Auto Code" for programming AI agents. The company says Auto Code matched the scores of Anthropic's "Claude Opus 5.5" on five benchmarks at 74% lower cost per task. Compared with the old version, however, the new default setting costs more per task. All figures come from the company's own evaluation, and it has not shown how they were aggregated or any third-party verification.

Google DeepMind on Nov. 6 (Nov. 7 Japan time) released research showing that AI agents can accelerate testing the strength of encryption. Under human guidance, the company's model found two weaknesses in legacy 1024-bit RSA, it said. According to media reports, signature forgery becomes about 1,600 times more efficient by a theoretical estimate, and batch factoring of keys cuts the average cost per additional key to about one-thirtieth. DeepMind said the current 2048-bit scheme is not affected by these weaknesses. The location of the paper and any third-party verification could not be confirmed.

Cursor, the AI coding tool, on Nov. 5 added a feature that lets users describe a tool they want to an AI agent, which then builds a web app, sets up a database, runs security checks, publishes it and allows it to be shared by link. It is available only with agents that run in the cloud, and it is offered on all paid plans with access to cloud agents. Nothing is published until the user chooses to, and at companies administrators can block publishing. Cursor has not disclosed hosting fees, usage limits or database capacity.

Cognition, maker of the AI programmer "Devin," on Nov. 5 announced "Devin Review 2.0," in which AI checks code changes. The company says it has cut the cost per review by 82% since August and made it about 1.5 times faster. A single review typically costs $0.25 to $0.75, a level at which it can be used on every change, the company says. All figures come from the company, no third-party verification could be confirmed, and it has not disclosed a false-positive rate. The pricing structure and which plans include it could not be confirmed either.

Perplexity on Nov. 5 announced that it had added "teammates," which keep working in set roles, to its AI agent "Computer." According to the company, they inherit Computer's projects, skills and memory, and each one runs on its own dedicated desktop and browser. They coordinate with other teammates and also contact colleagues on Slack. According to the product page, the feature is for subscribers to the paid "Max" plan and is offered first on the web. The size and duration of the free allowance during the promotional period and credit consumption after that are unconfirmed, and no outside verification of the isolation or performance was found. In always-on agents, OpenAI's "dots" and Anthropic's "Claude Duet" came first.

OpenAI on Nov. 3 (Nov. 4 Japan time) released "v0.22.0-alpha.1," an alpha version of its "Agents SDK" development kit for JavaScript and TypeScript, as a prerelease on GitHub. It embeds the runtime (harness) of the coding tool "Codex," which can handle model calls, tool execution, history and context summarization. It was introduced in a post by an outside developer, and no official OpenAI announcement had been found as of Nov. 5. The stable version on npm remains 0.21.0, and installation requires Node.js 24 and files distributed on GitHub. Existing in-progress state cannot be carried over. Pricing and the timing of a general release are unconfirmed.

OpenAI on Nov. 4 (Nov. 5 Japan time) made the ability to place phone calls from its voice AI "GPT-Live" generally available in its developer API. Realtime AI lead Justin Uberti announced it on X, and President Greg Brockman quoted the post. According to official documentation, outbound calling is available to all organizations, which can dial phone numbers through lines from providers such as Twilio and Telnyx and can also press buttons in automated phone menus. Audio costs $0.05 per minute, with carrier charges and fees for models used behind the scenes billed separately. Pricing specific to outbound calls, the regions covered and how consent and AI disclosure are handled could not be confirmed in the official documentation.

Charlie Holtz, founder of U.S.-based Conductor, which makes an app that runs multiple AI coding agents at the same time, said on X on Nov. 4 (Nov. 5 Japan time) that the company is joining Meta. According to Holtz, the team will build "the next interface" with Nat Friedman and Alexandr Wang at Meta Superintelligence Labs (MSL). Meta has not issued an announcement or comment. Axios reported, citing people familiar with the matter, that Meta acquired the company, but terms have not been disclosed. Conductor will shut down its cloud version on Nov. 20, and the version that runs on Macs will also wind down after updates through early 2027. On the same day, OpenAI also announced the acquisition of developer tool Pierre.

OpenAI on Nov. 4 published results from a system that uses AI to find vulnerabilities in the core of Linux, check whether attacks work and even produce proposed fixes. It said a single run using an unreleased cyber model validated 948 flaws that allow seizing administrator privileges, at an API cost of about $1.2 million. It also said it found at least 10 new types of attack techniques. All of these are the company's own statements; no third-party verification has been found, and they remain unconfirmed. The 948 is not the number of flaws already fixed, and the company has not disclosed the model's name.

Cat Wu, who served as product lead for Anthropic's AI coding tool "Claude Code," said on X on Nov. 4 (Nov. 5 Japan time) that she is joining OpenAI. She will lead product management for "Core Products," which includes ChatGPT and Codex. The announcement came five days after her last day at Anthropic; according to Business Insider, OpenAI confirmed the hire and Anthropic declined to comment. Her formal title, start date and scope of responsibility for Codex are unconfirmed. With Anthropic preparing for an IPO expected in the week of Nov. 16, the former head of one of its flagship products is moving to a rival.

OpenAI announced on Nov. 4 (Nov. 5 Japan time) that it will acquire Pierre Computer Company, a U.S. developer tools company. Pierre makes "Code Storage," which stores code in Git for AI agents, and open-source software that displays file lists and code diffs, the latter already used in the Codex app. OpenAI said it would make Codex "a place where teams can build software together with AI, from first idea to the version they ship." It has not disclosed the price or when the deal will close, which awaits regulatory approval. Axios reported that the deal targets Microsoft's GitHub. The acquisition follows OpenAI's purchases of Astral in March and Ona in June to strengthen Codex.

The official account of "Grok Bot," the AI agent of SpaceX subsidiary SpaceXAI (formerly xAI), announced on X on Nov. 4 (Nov. 5 Japan time) that a feature for making phone calls on users' behalf is now available. The feature was introduced by Elon Musk on Oct. 25, and in the promotional video, the bot gets the user's approval before dialing and the user can take over during the call. The post did not state which plans or countries are covered or whether the bot identifies itself as AI to the other party. According to a news report, the terms require it to identify itself as an assistant acting on the user's behalf, and it currently supports only U.S. numbers.

Game engine maker Unity on Nov. 3 posted a video on X introducing its official plugin for OpenAI's AI coding tool "Codex." Released in September, the plugin lets Codex choose and use on its own 31 kinds of task procedures (skills) written by Unity engineers. Unity has been extending the same skills to Anthropic's Claude Code, xAI's Grok Build, Google's Antigravity and GitHub Copilot CLI. The plugin does not connect directly to a running Unity Editor, and the version is at the beta stage. Unity will report July-September results on Nov. 5.

A research team from Tsinghua University and the University of California, Berkeley, on Nov. 1 published a paper on "uni-robot-agent," a framework that lets general-purpose AI models operate robots, and announced the results on X on Nov. 3. The code is released free of charge. Using Alibaba's open model "Qwen3.8-Flash-Next," the score on the robot manipulation benchmark "RoboDojo" rose from 1.90 with the previous framework to 31.39. The estimated cost per task was $1.09, far below $21.29 for OpenAI's GPT-6 Astra. Astra run on the same framework, however, scored 51.90 against 31.39, and 80.50 against 53.50 on 19 real-robot tasks, leaving a gap in performance.

OpenAI on Nov. 3 disclosed efforts to use ChatGPT and Codex for mechanical design, electrical design and simulation. Two employees used GPT-6 Astra to design an audio device for talking with Codex, from the enclosure to the circuit board and firmware, in less than a week, and a contract manufacturer built prototypes in less than a month. In preparing for mass production, however, the model could not consistently account for tolerances and assembly conditions, and human review was needed repeatedly. OpenAI is collaborating with Mazda on research into vehicle aerodynamic design and is testing with Chip Ganassi Racing the creation of drawings from 3D scans of parts. No new products or pricing were announced.

A video introducing new features of "Grok Bot," the AI agent from SpaceXAI (formerly xAI), was posted on X on Nov. 2, and Elon Musk also shared it as "Grok Bot improvements." It shows search results as product cards with photos and prices, places orders from product pages and tracks deliveries. Payment methods such as Shop Pay can be used, and each purchase requires the user's approval. It also added sharing of login credentials for bots used by teams and approval of connections through Slack.

AI coding tool maker Cursor on Oct. 27 added "Annotate" to the editor's "Design Mode." Users select components or draw circles and arrows by hand on the screen of a running web app to tell the AI agent where and how to make changes. Notes can also be collected and sent together. It is available in version 3.24.18 and later, but the documentation does not state which plans are eligible or whether there is an additional cost.

Cursor, the U.S. maker of an AI coding tool, announced on X on Oct. 26 the general availability of "Agent SDK," a development kit for defining custom AI agents as code. Instructions are written in Markdown and tools as TypeScript files, and after running evaluation tests, agents run on Cursor's platform or in users' own environments. Users can build bots that respond to GitHub pull requests or mentions in Slack. In every environment, however, model processing goes through Cursor, and usage fees are also paid to Cursor.

Cognition, the U.S. maker of the AI programmer "Devin," on Oct. 26 revamped "Devin Security," which handles everything from finding software vulnerabilities to prioritizing and fixing them. Using frontier models such as Anthropic's "Claude Mythos," a model for cyber defense, it searches large codebases even for attack paths that chain multiple flaws together. The company will offer enterprise customers hundreds of millions of dollars' worth of Devin usage credits. It has not, however, given any figures on accuracy or cost.

Researchers at China's Alibaba on Oct. 26 released the code, data and paper for "BabelArena," a benchmark that compares the abilities of AI agents across 23 languages. When five models from Anthropic, OpenAI, Google and Alibaba were tested, performance fell and tool-use errors increased in languages with fewer speakers. The researchers also showed an example in which Qwen-3.8-Max completed the same shopping task with 25 tool calls in English but took 201 in Telugu, a language of India.

"Managed Agents," a screen for managing AI agents in one place, is expected to be added soon to "Google AI Studio," Google's developer service, TestingCatalog, a U.S. outlet that tracks new features, reported on Oct. 25. It has five sections: agents, environments, credentials, triggers and run logs. Logan Kilpatrick, head of AI Studio, also mentioned a management screen built on top of Managed Agents. It would compete with Anthropic's platform of the same name. Google has not said when it will be available.

Elon Musk, who leads SpaceXAI (formerly xAI) under SpaceX, said on X on Oct. 25 that the AI agent "Grok Bot" can now call people and businesses on users' behalf. The bot states the user's request, takes notes on the other party's replies and sends a summary of the conversation. However, he did not disclose which plans or countries are eligible, what number is shown to the person receiving the call, or whether the bot identifies itself as an AI. Meta has also released an agent with a similar feature.

OpenAI updated the command-line version of its programming tool "Codex" to 0.164.0 on Oct. 23. The update added "presets" that let users save and recall up to 10 frequently used prompts, the ability to change the model and permissions together with a prompt, and sign-in to enterprise MCP servers. Thibault Sottiaux, who leads Codex, said use of the command-line version has doubled since the developer event "DevDay."

Cursor, maker of an AI programming tool, added "Computer Use" on Oct. 23, letting its AI agents operate desktop apps by clicking and typing. The agents can run tools such as the "iOS Simulator," used to test apps for Apple's iPhone, and test features they built themselves. Users can request it, and agents also use it on their own when needed. Running it on Macs managed by a company requires permission in the operating system.

Anthropic on Oct. 22 donated "Agent Skills," a system for teaching AI agents how to carry out work, to the Agentic AI Foundation (AAIF), an industry group under the Linux Foundation. According to the company, more than 70 agent platforms, including ChatGPT, Cursor and GitHub Copilot, have adopted it, and it has been installed about 280 million times. On the same day, OpenAI also donated "Open Responses," an API specification for calling different companies' models in the same format.

"Grok Bot," the AI agent developed by SpaceX subsidiary SpaceXAI (formerly xAI) and U.S.-based Cursor, added 6 features from Oct. 21 to 23. They are "Pages," in which the bot creates reports and dashboards and keeps them updated; sign-in on the user's behalf using the password manager 1Password; operation of phone apps such as the calendar and contacts; saving of logins and cards; video generation; and voice input that shows text as the user speaks. There is no setting that requires approval for each payment made with a saved card.

Anthropic on Oct. 21 began a research preview of "Claude Duet," a new app that alerts users to things they should pay attention to before they ask and keeps working while they are away. It regularly checks connected email and work services to create a morning summary, and continues tasks on a dedicated computer in the cloud. It always asks for confirmation before sending messages, making payments, signing or deleting. It is available to users of the paid "Pro" and "Max" plans in the United States by invitation from a waitlist. It shares usage limits with regular Claude.

OpenAI said on Oct. 20 that it had acquired U.S.-based Orchestra Bio, which makes software for managing the progress of drug development programs. It will be folded into "Rosalind," its life sciences business, expanding from laboratory work to coordinating trials, contractors and schedules. The purchase price was not disclosed. The next day, Oct. 21, the company published research in which thousands of AIs running Rosalind's model "GPT-Rosalind" cross-checked large amounts of mouse brain cell data against scientific papers in 32 minutes, pointing to clues about a rare neurological disease. However, the paper has not been peer-reviewed, and the hypothesis has not been tested experimentally.

Two serious flaws that let attackers run arbitrary programs remotely were found in PHP, a language widely used for website programs, and an emergency fix was released on Oct. 21. OpenAI President Greg Brockman said the flaws were found using the company's model and that it spent two months forming a coalition with the companies involved to address them. The flaws are said to have existed for 14 years. Finland, the Netherlands and France temporarily took government websites offline the same day.

Thibault Sottiaux, who leads OpenAI's programming assistant "Codex," said on Oct. 20 that, after reports over the weekend that Codex usage allowances were running down faster than expected, about 40,000 accounts had been assigned usage limits that differed from the plans they subscribe to. The bug has been fixed, and affected users were contacted individually and given usage credits as compensation. He did not say which plans were affected or in which direction the limits were off.

OpenAI on Oct. 20 added an option in its API to keep "prompt caching," which temporarily stores identical input sent repeatedly and reads it back at a lower price, for 12 hours. The standard had been 30 minutes, so for always-on agents that run every few hours the cache expired and the discounted pricing could not be used. Writes to the 12-hour cache cost 2.5 times standard input, while reads cost one-tenth of standard. Whether it saves money depends on how much identical input is reused.

Perplexity on Oct. 20 made "Computer," its AI agent that works on a computer on the user's behalf, able to run entirely on a Mac. The model, task planning and file processing run on the device, and it operates apps such as Notes, Mail and Keynote. It is available to paid "Pro" and "Max" subscribers and business customers, and 36GB or more of memory is recommended. Searching for the latest information and operating a browser require the cloud, and the user grants permission each time. The model used on the device is based on "Qwen" from China's Alibaba.

Matt Palmer, who works on "Grok Bot," the AI agent from SpaceX subsidiary SpaceXAI (formerly xAI), presented 12 updates to the bot in a video on Oct. 20. The bot can now read and reply to direct messages (DMs) on X, and it handles requests when called with "@bot" in an X post. On Android it can be set as the default assistant, and work on the bot's computer has become about three times faster on average. The company did not say whether DM replies require approval each time.

"Grok Bot," the AI agent from SpaceXAI (formerly xAI), a SpaceX subsidiary, became assignable to the "Action button" on the side of the iPhone on Oct. 19. Pressing the button starts a voice conversation with the Bot, like a phone call. It works on iPhone 15 Pro and later models, which have the Action button. Grok Bot gives each user a computer in the cloud and keeps working after a conversation ends, and it competes with OpenAI's always-on agent "dots."

Open-source AI coding tool "OpenCode" disclosed on Oct. 17 that API keys from OpenAI, Anthropic and Google that users had registered with its paid management service, "OpenCode Console," may have leaked. An old system slated for retirement was accessed without authorization, affecting 1.4% of Console users. The company said the incident was linked to a large-scale distillation attack, but has not disclosed how the intrusion occurred or how many keys leaked. Users have reported a string of unexplained charges ranging from several hundred dollars to $1,700.

Elon Musk, who leads SpaceX subsidiary SpaceXAI (formerly xAI), posted on X on Nov. 1 (Nov. 2 Japan time) that the cost of serving "Grok Bot," an AI agent that carries out work on users' behalf, was unsustainably high relative to revenue. He was replying to a user who asked for a reset after their weekly allowance ran out with several days left. He did not say how the limits had been changed, and the size of the allowance has not been disclosed. Two weeks earlier, at an earnings call, the company said paid weekly users had reached 800,000. In the same week, OpenAI and Anthropic expanded the limits on their paid plans.

Microsoft on Oct. 30 released on Hugging Face the weights of "Mage-VL-Cowork," an AI model that handles computer work by switching between operating the screen and running commands within a single model. It is small, at about 4 billion parameters, and in the company's evaluation it succeeded on 66.0% of tasks on the computer-use benchmark "OSWorld-Verified," above 30-billion-parameter-class open models. However, it fell short of top commercial models (GPT-5.5 at 78.7%) and the human baseline (72.4%), and the figures have not been verified by third parties. The license is Apache 2.0, which permits commercial use, and Microsoft has not announced the model on its blog or elsewhere.

The British government said on Oct. 30 (Oct. 31 Japan time) that "PlanAssist," an AI developed with Google DeepMind and others to support the assessment of housing planning applications, will expand from a trial at 3 councils to 15 councils and be used on real applications for 12 months. Google DeepMind said on Oct. 31 that with the AI, built with Gemini, the time officers spent assessing householder applications fell by an average of 58%, and more than 70% of the reports it produced were approved without changes. The government aims to make it available to all councils in England by the end of 2027. However, the figures come only from Google and government announcements, the measurement method has not been disclosed, and the metric differs from the goal of halving the time to a decision set in June.

OpenAI said on Oct. 30 (Oct. 31 Japan time) that it had added five features to its "Agents API," which lets companies build AI agents into their own products. The centerpiece is "Auto Review," in which a separate AI reviews actions an agent is about to take and stops dangerous ones, a mechanism ported from its coding assistant "Codex." It also added a feature that summarizes context without halting long-running work and an experimental feature that tells the AI how much time has passed. Auto Review is disabled by default, and OpenAI has not published a list of the actions subject to review. The Agents API began a public beta in September, and Anthropic and others offer similar platforms.