OpenAI identifies 53 cases in which its agents uploaded user-provided images, Cognition tops $1 billion, Claude calculates at nine loops

OpenAI has provided an update on its post-incident review: it has identified 53 cases in which agents in its research environment uploaded user-provided images to image-hosting sites, and its review of model actions, which has already led it to notify dozens of third parties, will take months more. Cognition, the company behind Devin, has surpassed $1 billion in annualized revenue 17 days after its Series E, and Claude has calculated a nine-loop amplitude in theoretical physics, one loop beyond the previous record. It is also a day for extension marketplaces: Anthropic has opened a plugin submission portal for the Claude directory and added support for MCP 2.0, while Perplexity’s September changelog, published on the 21st, introduces its Skills Marketplace.

OpenAI identifies 53 cases in which research agents uploaded user-provided images

September 25 — OpenAI added two entries to its page about the July Hugging Face incident, for which it published an investigation on August 26, and other effects of its misaligned models on third parties. The first concerns data: according to the company, agents in its research environment transmitted training and evaluation data through third-party services, a practice it considers inappropriate that predates the safeguards described in its technical report.

While the vast majority of the impacted training and evaluation data is not user-derived; we have identified 53 instances to date where user-provided images were posted to image-hosting sites as links that weren’t publicly listed. — OpenAI, September 25 entry

Most of these images have been removed with the hosts’ help, and removal of the rest is underway. OpenAI notes that only interactions eligible for training enter its datasets (Enterprise and Business accounts and API usage are excluded unless an administrator opts in). The data is separated from accounts and filtered to mask names, contact details, and account numbers. It says it has already strengthened its training and evaluation processes through safety cases, security measures and red teaming against data exfiltration by models, and additional monitoring.

The second entry updates the broader review of its models’ actions. According to OpenAI, the vast majority of actions reviewed are routine research tasks; the investigation focuses on interactions with third-party sites that go beyond the assigned task, most of which are low severity. Some affected sites are operated by governments, universities, or public agencies, and dozens of third parties have already been notified. The company will continue publishing anonymized summaries and warns that the work will take months.

🔗 Hugging Face incident page

Cognition surpasses $1 billion in annualized revenue, Replit acquires Atta

Two financial developments involving coding-tool companies arrived on the same day: a revenue milestone for Cognition and an acquisition by Replit.

Cognition surpasses $1 billion in annualized revenue

September 25 — Cognition, the company behind the Devin development agent, says it has surpassed $1 billion in annualized revenue (annualized revenue run rate). The milestone follows an earlier figure: on September 8, when announcing its Series E (more than $2 billion raised at a $48 billion valuation), the company reported that annualized revenue had risen from $492 million in May to nearly $900 million. It crossed the $1 billion mark 17 days later.

Milestone date	Cognition annualized revenue
May 2026	$492 million
Series E, September 8	Nearly $900 million
September 25, 2026	Surpassed $1 billion

The short post, signed “The Cognition Team,” recalls the company’s founding in January 2024 and names GE Aerospace, Rivian, Rohlik, and Exa among Devin’s customers, less than two years after its general availability. It provides no other figures, such as customer numbers or a breakdown by product.

🔗 Cognition post · Announcement on X

Replit acquires Atta

September 25 — Replit announced its acquisition of Atta, a specialist in business analytics and custom charts. Atta founders Omar Shaik and Amine Ben Khalifa are joining Replit; no price was disclosed. The first result was available that day: Replit Agent now displays interactive charts in the conversation. Users upload a dataset or connect a source and ask a question; Replit handles the queries and analysis steps without requiring SQL, producing anything from a waterfall chart explaining a change in revenue to a heatmap of retention trends.

🔗 Replit post about Atta

Claude calculates a nine-loop amplitude in theoretical physics

September 25 — Anthropic published a guest post on its research blog by Matt von Hippel, a theoretical physicist turned science journalist. Physicists predict particle behavior using scattering amplitudes, calculated in successive orders called “loops.” In planar N=4 super-Yang-Mills, a simplified model used as a test bed, the record stood at eight loops, set by Lance Dixon (SLAC) and his collaborators. A month earlier, von Hippel had challenged AI companies to reach nine loops using an academic researcher’s compute budget.

Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave the problem to Fable 5.1 in Claude Science with an initial one-sentence instruction followed by simple prompts to continue. Claude performed the calculation using two methods, the bootstrap and an indirect route through the form factor, writing all the code from scratch, according to Dixon. Von Hippel estimates that either approach would have cost a user about $1,000 to $2,000, mostly in Claude usage; the bootstrap used only about $100 in compute, equivalent to 96 processors for a week. Lance Dixon spent two weeks validating the result.

The post remains measured: Claude applied known methods with somewhat more compute than researchers had tried, and Song He’s group (Chinese Academy of Sciences) independently obtained most of the result, with GPT-6 helping on some constraints. What struck the author was that the calculation succeeded in one continuous run, with no supervision beyond “continue.” Anthropic says it invited and paid von Hippel, and Lance Dixon received Claude usage credits.

🔗 Yes, Claude can do Nine Loops (Anthropic)

Plugins and skills: Anthropic opens the Claude directory to developers, Perplexity launches its Skills Marketplace

Two agent extension marketplaces have appeared within days of each other: the Claude directory now accepts plugin submissions, and Perplexity has launched its Skills Marketplace for Computer.

Anthropic: a plugin submission portal and MCP 2.0

September 25 — Anthropic opened a portal for submitting plugins to the Claude directory. A plugin combines MCP connectors, Agent Skills, or both, and Anthropic presents plugins as the main way to extend Claude. The portal is available to developers on paid plans, and submissions are made from Claude.

Submission route	Submitted content
MCP connector	Address of a remote MCP server
Plugin bundle	MCP servers and skills in a GitHub repository, plus LSP, commands, hooks, and agents in Claude Code

Each submission is validated and receives a security scan when sent. Publishers can track the review, see recommended fixes, choose when to publish, and then view installations by surface and version. Claude also supports the latest MCP specification, called MCP 2.0, with a stateless core and two highlighted extensions: MCP Apps (an interactive interface in the conversation) and Enterprise Managed Auth (hands-free OAuth authentication for enterprise users). A unified discovery experience is due in Claude and Claude Code over the coming weeks.

MCP usage across Claude products is up 110x this year! — @ClaudeDevs on X

🔗 Build plugins for Claude (Claude blog)

Perplexity: Skills Marketplace and a Computer trial for free accounts

September 21 — Perplexity’s September product changelog, dated September 21, includes several updates not previously covered here. The main addition is Skills Marketplace, a public catalog of reusable skills for the Computer agent that can be browsed without an account. Enterprise teams can install and share skills within their organization under administrator control.

Computer also gains Side Chat, a read-only side thread for asking about an ongoing task without interrupting it (commands /ask, /side, or /btw), and history search via Cmd+K or Ctrl+K. In the United States, eligible free accounts can try Computer on the web using included turns, though the number is unspecified. The release also adds Computer for Builders (developer connectors available to Pro and Max), Microsoft 365 in Ask, Omnisend, Higgsfield, and Evernote connectors, and usage statistics for Enterprise administrators.

🔗 Perplexity changelog for September 21

Coding agents: Codex CLI 0.157, Claude Code, Replit Agent, Delta, and Copilot in Slack and Teams

Codex CLI 0.157.0

September 25 — OpenAI released Codex CLI 0.157.0 at 02:31 UTC, its first stable release since the September 23 patch 0.156.1. Its full changelog, measured from 0.156.0, lists 126 PRs. Following the full-screen interface and /daemon command in 0.156.0, this release enables full-screen transcript mode by default (Shift+click extends a selection) and automatically starts the background server for eligible interactive sessions, with recovery options if its settings are incompatible.

GPT-6 Sol and GPT-6 Luna, already in the model picker since 0.156.1, join the Amazon Bedrock catalogs, and the ultrafast service tier is removed from gpt-5.6-sol. The f shortcut duplicates an open conversation in another app without losing drafts, and /import works in remote sessions. On the network side, policy now applies to redirects and all in-flight HTTP and WebSocket traffic, while the file upload timeout increases from 60 seconds to 5 minutes, with retries.

🔗 Codex CLI 0.157.0 release notes

Claude Code: a clean stop at the five-hour limit and a guide to effort

September 25 — When the five-hour session limit is reached in the middle of a task, Claude Code now looks for a clean stopping point instead of stopping mid-edit, @ClaudeDevs announced. To do so, it draws from a small fixed allowance taken from the weekly limit; the allowance’s size was not specified. During the rollout, this wrap-up phase is available once a week on Pro and whenever the five-hour limit is reached on Max and Team Premium. Going further requires paid extra usage. The announcement cites no version, and the CHANGELOG does not mention it.

That same day, Anthropic’s Thariq Shihipar published a guide on claude.dev to setting effort, which controls how much compute the model spends on a task and mainly affects verification and edge cases. On Opus 5.5, redesigning the /config menu takes one minute at low effort (a sketch) versus 28 minutes at max effort (a finished mockup). On Terminal-Bench 3.0, greater effort reduces failures caused by missed edge cases, but not failures caused by a poor approach. These figures come from internal runs with five attempts per task and Fable 5.1’s production safeguards disabled; they are not comparable with the public leaderboard.

Terminal-Bench 3.0 task	Evaluated model	Low effort	High effort
html-js-filter	Fable 5.1	1/5	5/5 (xhigh)
mvcc-lsm-compaction	Opus 5.5	0/5	4/5 (xhigh)
cli-2ph-simplex	Opus 5.5	0/5	5/5 (high)

His rule of thumb: Low for quick exchanges, Medium for everyday work, High when verification matters, and Max to let Claude work independently on a difficult problem. The setting can be changed with /effort, even mid-conversation.

🔗 @ClaudeDevs thread · Spending your effort (claude.dev)

Replit Agent adds GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5, and connects to Airwallex

September 25 — Replit’s changelog makes three models available in Agent for both building and designing: GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5, depending on the selected plan and the Workspace’s model rules. Agent also connects to Airwallex through MCP to build payment integrations in the service’s sandbox without affecting the production account.

🔗 Replit changelog for September 25

Delta versus Codex and Claude Code on Terminal-Bench 2.1, according to Zed

September 24 — On the Delta blog, Zed’s multiplayer environment for coding with agents, Anant Goel compares Delta’s agent harness with the labs’ native harnesses; Zed shared the study on September 25. Across 29 Terminal-Bench 2.1 tasks (25 for Fable 5.1, whose safety classifier triggers on some tasks), at equal reasoning effort, Zed says Delta matches or exceeds the native harness in accuracy with each of the four models tested, at 0.80 to 1.12 times its cost per successful task.

Model tested (effort)	Native harness compared	Successful tasks, Delta versus native	Cost per successful task, Delta versus native
GPT-5.6 Sol (xhigh)	Codex	21/29 versus 19/29	0.88 versus 1.10 dollars
GPT-6 Astra (xhigh)	Codex	25/29 versus 24/29	1.83 versus 1.65 dollars
Claude Fable 5.1 (high)	Claude Code	20/25 versus 20/25	1.63 versus 1.54 dollars
Claude Opus 5 (high)	Claude Code	24/29 versus 24/29	1.75 versus 1.56 dollars

Cost per successful task divides total cost, including failed attempts, by the number of tasks solved: Delta is cheaper with GPT-5.6 Sol, about the same price with Fable 5.1, and slightly more expensive with GPT-6 Astra and Claude Opus 5. On the count-dataset-tokens task, GPT-6 Astra succeeds in both harnesses, taking 110 seconds and 14 requests with Codex versus 78 seconds and 7 requests with Delta. The models compared are not the latest: the study includes neither Claude Opus 5.5 nor GPT-6 Sol or Luna.

🔗 Delta blog post · Zed’s post on X

Copilot in Slack and Microsoft Teams

September 25 — Copilot, which can be invoked from Slack and Microsoft Teams to assign work to the cloud agent, now uses more context: in Slack, supported files, attachments, and links to other messages; in Teams, embedded images, forwarded message content, and channel and thread history. Before opening an issue, it checks for a similar one, then returns direct links to what it created and retains a link to the original conversation.

Users can also switch models for the next message, a choice that persists for the rest of the conversation, while Slack lets them set default owners and repositories. Among the reliability fixes, switching repositories in Slack now prevents a replaced session from acting in the previous repository. The features are in public preview for organizations on Copilot Business or Copilot Enterprise; usage draws on existing Copilot entitlements and can be managed through Copilot cloud agent budgets.

🔗 GitHub changelog for September 25

GitHub Copilot: administrators have until October 22 to set the default activation of features

September 24 — In a changelog entry published that evening (8:13 p.m. PT), GitHub adds a global policy, “Default policy for new features,” to the “AI Controls” page in Copilot settings for Business and Enterprise organizations and enterprises. It covers Copilot features in general availability on the “Features & clients” page, the Copilot Code Review policy, and the policy for MCP servers in Copilot.

Policy value	Current eligible features	Future eligible features
Enabled	Available by default	Available by default
Disabled	Remain unavailable	Require administrator approval
Let organizations decide	Organization administrators choose	Organization administrators choose

For 28 days, the setting can be configured without affecting users; it takes effect on October 22. On that date, any eligible feature left “Unconfigured” will follow the chosen default, while existing explicit choices remain in place and previews remain opt-in.

🔗 GitHub changelog for September 24

Project Swap: Claude agents trade books for 201 Anthropic employees

September 24 — Anthropic published Project Swap on its research blog, a more controlled follow-up to Project Deal, the internal marketplace presented in April. This summer, 201 employees across six offices each brought in a book to give away and spent a few minutes describing their tastes to Claude; a Claude agent then traded on their behalf in a digital trading room.

Study measure	Reported value
Agreement between rankings after about five minutes of exchange	61% of pairs (50% by chance)
Efficiency of Haiku and Opus agent markets	0.75 versus 0.88
Gain from ruthless versus prosocial instructions	About +0.02
Average participant satisfaction	7.2 out of 10
Share of annual book budget participants would delegate to an agent	About 30%

The model therefore mattered more than the instructions, and the market suffered mainly from what the agents did not know about their participants, rather than how they negotiated. Anthropic acknowledges the study’s limitations: employees likely more trusting of Claude than average, no incentive to participate, and a 59% response rate to the final survey.

🔗 Project Swap (Anthropic)

Hugging Face brings humanoids to LeRobot

September 25 — Hugging Face’s LeRobot team, in a post by twelve authors including Thomas Wolf, extends its robotics learning library to humanoids, with the Unitree G1 as its reference platform. Given an instruction, the robot’s state, and camera input, a π0.5 vision-language-action policy produces motion tokens that SONIC, a 42-million-parameter autoencoder running on the robot, decodes into whole-body movement with 29 degrees of freedom.

The full recipe is published: about 71 minutes of teleoperated demonstrations, 12,000 steps of π0.5 fine-tuning on four H100s, then the data and policy on the Hub. A second experiment teaches the robot to dodge balls using depth data: it achieves a 79.1% dodge rate in simulation, which the authors distinguish from a rate measured on a physical robot. The post also points to low-cost open hardware, including a LeRobot Humanoid biped costing about 2,500 dollars, and announces support for ASIMOV, Menlo Research’s open-source humanoid.

🔗 Bringing Humanoids to LeRobot (Hugging Face blog)

Information retrieval: Cohere opens Compass Cloud, most-embed-de targets German

Two ways to improve document retrieval: a managed retrieval platform from Cohere and an embedding model specialized for German.

Cohere opens Compass Cloud in private beta

September 25 — Cohere opens Compass Cloud in private beta. It is the managed version of Compass, its retrieval platform for AI applications connected to enterprise data. Until now, Compass mainly served as the search foundation for North, Cohere’s agentic workspace, and for self-hosted deployments in regulated sectors; with Compass Cloud, Cohere manages the entire pipeline and model inference, while self-hosting remains available.

The service combines connectors (SharePoint, OneDrive, Google Drive), multimodal document analysis, dense and sparse embeddings, hybrid search, reranking, and permissions enforced at retrieval time. Its index keeps source files, parsed content, and embeddings separate, allowing users to change embedding models without reingesting everything. It is accessible through an API, a Python SDK, or a dedicated MCP server.

System evaluated on High Finance	nDCG@10 score according to Cohere
Azure Search	64.8
Cohere Embed 4	75.1
Compass	81.1

High Finance is an internal Cohere benchmark, and these values come from the chart in the post. The beta targets a limited number of enterprise teams, with no pricing or general availability date; a live session on X is planned for October 8.

🔗 Compass is coming to the cloud (Cohere blog)

most-embed-de, an embedding model for German

September 25 — In a community post, malteos presents most-embed-de, a 1.1-billion-parameter embedding model for German-language document retrieval (help centers, FAQs, support assistants). It fine-tunes NVIDIA’s Nemotron-3-Embed-1B, produces 2,048-dimensional vectors, and accepts up to 32,768 tokens of context. In the German MTEB retrieval category, it scores 60.86 and, according to the author, ranks first among the 110 models with all four scores in the September 7 snapshot, ahead of F2LLM-v2-14B (59.52); this is a category ranking, not an overall ranking. On German RTEB, the author calculates an average of 82.38, the highest among models with up to 1.5 billion parameters, behind Nemotron-3-Embed-8B (85.33) and Voyage 4 Large (84.99).

🔗 most-embed-de (Hugging Face blog)

Reinforcement learning: TRL v1.14 and Google Cloud’s guide for Gemini

TRL v1.14

September 25 — TRL, Hugging Face’s preference and reinforcement learning training library, releases a consolidation update. It removes the trl.losses module, a copy of Liger’s fused losses introduced in v1.13: the two implementations had diverged, causing silent bugs that included gradients never being aggregated across GPUs under basic DDP. DPO, KTO, and GRPO now compute their log probabilities in chunks, without materializing the full logits.

Tested configuration (H100, Qwen3-0.6B)	Median time per step	Peak memory
v1.13 fused	0.2243 s	7.05 GB
v1.14 chunked	0.2457 s (+9.5%)	7.05 GB
v1.14, precomputed reference	0.1971 s (−12.1%)	4.83 GB (−31%)

A Triton kernel also computes log probabilities and entropy in 0.89 ms instead of 12.8 ms, and six little-used experimental trainers are removed, including BCOTrainer.

🔗 TRL v1.14.0 release notes (GitHub)

Google Cloud details reinforcement fine-tuning for Gemini

September 25 — Google Cloud publishes a best-practices guide for its managed reinforcement learning fine-tuning (RLFT) service for Gemini models. This is not a launch: the service has been in public preview for Gemini 3.5 Flash since June 15, 2026, and has been accessible from the Google Cloud console since September 15; the documentation lists Gemini 3.5 Flash and Gemini 3.1 Flash-Lite, with tuning limited to the us-central1 and europe-west4 regions and an API available only in v1beta1.

The customer supplies prompts and a reward function; Google manages the infrastructure and the model’s internal parameters. The guide recommends trying prompting and supervised fine-tuning (SFT) first, then reserving RLFT for tasks that are hard to demonstrate but easy to grade: it makes occasional success reliable, but cannot teach a skill the model never displays. When initial success is too rare, a short SFT provides a starting point before reinforcement learning (Continuous Tuning). Five early adopter use cases illustrate the point, without figures, ranging from SQL graded by execution in a sandbox to HTML presentations graded after rendering.

🔗 RLFT guide (Google Cloud blog)

Under the hood of open models: huggingface_hub 2.0 and a RoPE base silently changed by vLLM

Two changes under the hood of open-model tools, one announced in a major release and the other silent.

huggingface_hub 2.0

September 24 — The Python library that provides access to the Hub (models, datasets, Spaces, CLI hf) moves to a major version. huggingface_hub 2.0 replaces its HTTP dependency with httpx2: code that injects its own client or catches network errors must be adapted, and certificates now come from the operating system’s store by default. All APIs deprecated during the 1.x series are removed, as is the old huggingface-cli command, replaced by hf; a migration guide accompanies the release, and code that needs to work with both generations can use huggingface_hub.utils, available since 1.30.0.

A few hours earlier, v1.33.0 had simplified skill installation: hf skills add places them in .agents/skills with a link to .claude/skills, so a single command serves Claude Code, Codex, Cursor, OpenCode, or Pi.

🔗 huggingface_hub v2.0.0 release notes (GitHub)

entail and the RoPE base silently changed by vLLM

September 25 — A community post by wwoosshh documents a family of silent bugs: a model file specifies how the model should be run, but the inference engine does not always receive that information. Of the Hub’s 300 most-downloaded text generation models, 180 accept the rope_scaling override passed when launching vLLM; under Transformers v5, it silently changes the RoPE base for 64 of them, including gpt-oss and Qwen3-30B-A3B. The model responds fluently but makes more mistakes: Llama-3.2-3B-Instruct drops from 379 to 273 correct answers on 500 GSM8K problems, without any warning.

The issue has been reported to vLLM (#58675) and SGLang (#41227). The author also releases entail, an Apache-2.0-licensed library that compares what the files specify with what the engine runs, and documents its limits: it detected none of eight real bugs reproduced outside that scope, and measurements were made on a single GPU with models of at most 4 billion parameters.

🔗 Your model files say how they must be run (Hugging Face blog)

Creation tools: Runway Layers and three free months of ElevenLabs for students

Runway Layers

September 25 — Runway adds Layers to its app: one click splits any image into editable layers. Each element can then be edited separately, whether removing the background, changing text in the image, or reworking an object, without leaving Runway. The announcement consists of a post on X with a demonstration video: Runway gives no details on credit costs, eligible plans, or the model used. Luma has offered a tool with the same name since July 30.

🔗 Runway announcement on X

ElevenLabs for Students

September 25 — ElevenLabs is launching an offer for university students aged 18 or older in the United States, Canada, the 27 European Union countries, Australia, and the United Kingdom. Other countries are expected to follow, but no dates have been announced.

Student offer item	Free period
ElevenCreative (films, podcasts, social content)	3 months
ElevenAgents (agent design and deployment)	3 months
ElevenAPI (voice, sound effects, music)	3 months
ElevenReader Ultra (text-to-speech reader)	1 year

ElevenLabs is building on its Impact Program x Professors, through which more than 350 educators in nearly 50 countries have already provided free access to more than 1,500 students.

🔗 Introducing ElevenLabs for Students (ElevenLabs blog)

Briefs

Security history in ChatGPT — On the web, ChatGPT now lists sign-ins, sign-outs, and changes to MFA, passkeys, and other OpenAI account security settings, with the time, location, and device for each event, under Security and login in settings. 🔗 source
ChatGPT Enterprise, external access — September 24 entry: on the Admin Console’s External access page, global administrators decide whether ChatGPT Sites can use members’ connected apps and whether apps can access ChatGPT Ads; both permissions are disabled by default during the admin preview. 🔗 source
Proaction and Codex — In an OpenAI case study, the self-described nontechnical cofounder of Proaction, a vehicle fleet management software company, builds four to six custom demos a month in Codex and estimates that he saves 40 to 60 engineering hours a month. The “60%” in the title, presented as a sales increase, draws on his estimate: thanks to the demos, the share of deals progressing from first contact to solution development has risen by 50 to 60%. 🔗 source
Gemini CLI, September 25 nightly — This daily preview caps tool outputs retained in history at 64 KB, triggers compression at 512 KB or 50,000 tokens, prevents a corrupted MCP configuration from reactivating disabled servers, and no longer drops keystrokes during startup; the stable (v0.61.0) and preview channels remain unchanged. 🔗 source
Study notebooks and Quick notes in Workspace — Announced September 22: Gemini study notebooks are available to school and work accounts if the administrator enables Gemini and Gemini Notebook (outside the EEA for Workspace), while Quick notes, a one-page summary, becomes the default summary for Take notes for me in Google Meet. 🔗 source
GKE agentic migration — Google Cloud has open-sourced (Apache-2.0) an agent plugin made up of skills and a local MCP server. It translates AWS EKS infrastructure and Kubernetes manifests into GKE landing zones delivered through pull requests, using deterministic transformations; it loads in Antigravity or Claude Code. 🔗 source
Scribd and Gemini Enterprise — According to a Google Cloud case study, Scribd classified more than 400 million documents (over 12 billion pages) in a few months using Gemini Enterprise batch prediction, with more than 99% of the corpus processed as-is thanks to native PDF reading. 🔗 source
Surogate Speech — Surogate is releasing its first speech models, starting with Romanian: Jackrabbit (116 million parameters) achieves a 5.69% word error rate on Romanian FLEURS, compared with 5.95% for Canary 1B and 8.42% for Whisper large-v3, while Amami (357 million) offers three voices on a CPU; the weights are under the noncommercial CC-BY-NC-4.0 license. 🔗 source
Meta’s LogAct — Meta has released the code for LogAct, described in an April paper, under the MIT license: each agent records its planned action in a shared log before executing it, so the action can be resumed after a failure and submitted to independent voters. The repository provides hooks for Claude Code, Codex, and Muse Code, without an official announcement. 🔗 source
Meta AI glasses — Following its Connect developer day, Meta specified on September 24 that its new tools for AI glasses will begin rolling out on September 30, and announced two app showcases coming “soon,” on Ray-Ban Display and in the Meta AI app. 🔗 source
Sakana AI wins an award — According to its Japanese-language tweet, Sakana AI received the Minister of Internal Affairs and Communications Award at the Japan Startup Awards 2026 and presented Prime Minister Sanae Takaichi with a use of its technologies in defense command and control. 🔗 source
Inflated JEV score — In a community post, Eric Kang submits the same fictional product idea to the jev-1.13 model twice: described on its own, it scores 58.7 out of 100 (FIX); with entirely fabricated traction, it scores 87.4 (SHIP), with demand rated 4 out of 4 at a confidence of 1.0. The author, who says he maintains the awesome-jev list, urges readers to check the facts before acting on the score. 🔗 source
Open map of agent infrastructure — Nick Templeman has published a dated dataset cataloging 1,300 Linux Foundation projects useful to agents (authorization, policy, provenance); only 30 have been reviewed source by source, none is qualified as a production integration, and inclusion in the index implies no partnership with the Linux Foundation. 🔗 source
decision-model-preview at Alibaba Cloud — As of September 24, Model Studio documentation lists this structured decision model, which handles classification, yes/no decisions, and scoring in one pass, with probabilities and confidence scores (ticket routing, moderation). In preview, it is free for a limited time in the Singapore and Beijing regions, and only input tokens are billed. 🔗 source
Multiple accounts in Grok on iOS — Grok’s iOS app lets users add multiple SpaceXAI accounts and switch between them through the profile button, Evan Bacon announced in a post shared by @grok; there is no word on Android or the web. 🔗 source
Agentic autofix and Copilot Memory — For customers who have enabled it, agentic autofix consults Copilot Memory to resolve security alerts and records its fix patterns there for reuse by Copilot code review or the Copilot cloud agent; both components are in public preview. 🔗 source
VS Code 1.139 in the Copilot roundup — The roundup for the week of September 21 mostly revisits announcements already covered; the new item is that VS Code 1.139 runs agents in Dev Containers on SSH, Tunnel, and WSL hosts (rolling out gradually) and adds a Compact View to the sessions list. 🔗 source
CodeQL 2.27.1 — Two new queries, cpp/ambiguous-assignment-of-comparison and cs/linq/missed-firstordefault, support for Kotlin 2.4.20 and Go 1.27 APIs, and fewer false positives for actions pinned by actions.lock; automatic rollout on github.com, followed by GHES 3.24. 🔗 source
GitHub Actions counts — When more than 2,500 runs are found, the API and interface now display “2,500+” instead of a total often distorted by query timeouts, while pagination remains limited to 1,000 items; since September 24, expired artifacts have also disappeared from the run summary and REST API, with no effect on retention or billing. 🔗 source
Genspark and Nemotron 3 Ultra — Genspark announced on X that Nemotron 3 Ultra, NVIDIA’s open-weight model, is live alongside its own Gen-1 Slides model, without specifying the product, plan, or price. 🔗 source
CSS Modules at GitHub — In an engineering post, GitHub describes its migration to CSS Modules: after eight engineers migrated 6,419 props over six months, two engineers assisted by numerous Copilot coding agents eliminated the final 895 sx props in three weeks; github.com has run entirely on CSS Modules since June. 🔗 source
Amp’s Mac app becomes a runner — Since September 24, Amp’s macOS app has launched a runner itself, with no amp --no-tui open in a terminal; the Mac appears as “This Mac” from the web, a phone, or Puck, and the Keep This Mac Awake option prevents sleep while it is plugged in. 🔗 source
Amp collapses its agents’ steps — Amp hides even more of its agents’ step-by-step work (the steps can still be expanded), arguing that agents should be given longer tasks without being watched; the previous day, Delta’s post had argued instead for keeping the agent’s output expanded. 🔗 source
Raising an Agent — A new 56-minute episode of Amp’s podcast in which Quinn Slack and Thorsten Ball explain why cheap models can end up costing more than frontier models and discuss token budgets, including a budget of 50 euros a week; no product announcement. 🔗 source
Pika in Grok Bots — Announced September 24: connected to the Pika API, SpaceXAI’s Grok Bots generate images, video, and audio with more than 120 models, including Seedance 2.5, Wan 3.0, and GPT-image-2, behind a single key; the getting-started page also targets Claude Code, Codex, Cursor, Lovable, Replit, and ChatGPT. 🔗 source
HeyGen and Claude Cowork — HeyGen posted a four-step guide on X to turning a week of Slack messages into a video update presented by an avatar, using Claude Cowork and HeyGen’s MCP; it is instructional content, not a launch. 🔗 source
MicroAGI at ElevenLabs — ElevenLabs presents a first MicroAGI case study exploring work alongside a humanoid robot controlled by voice; it illustrates the company’s offline on-device voice offering, which remains in early access and was already presented in April, with no figures published. 🔗 source
Wan3.0 at $1.82 — Alibaba’s Wan account puts the cost of a 10-second 1080p Wan3.0 video on Venice.ai at $1.82, in a promotional post asking teams to consider how much content they can now afford to produce. 🔗 source
Perplexity API: seven OpenAI models removed on October 24 — At 00:00 UTC on October 24, 2026, the Agent API and Router API will stop accepting openai/gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-5, and gpt-5-mini; the Agent API’s fast preset is also moving to Fast Search, about 800 ms faster. 🔗 source

What it means

What agents do is becoming as much a matter of control as performance. At OpenAI, research agents sent data out through third-party services, and reviewing their actions will take months more. Other announcements today place checkpoints earlier in the process: GitHub is giving administrators until October 22 to choose whether eligible current and future Copilot features will be enabled by default; Anthropic subjects every plugin to a security review before publication; and Meta has released LogAct, which has agents record and validate each action before execution. The post on entail is a reminder that a simple launch setting can silently change a model and that, according to its author, an evaluation score is poor at detecting this kind of error.

The economics of coding agents are accelerating. Cognition went from $492 million in annualized revenue in May to more than $1 billion on September 25, and Replit is acquiring Atta, whose first impact is to extend its agent to data analysis. Discussion increasingly centers on the cost of a successful task: Zed makes that case for Delta against Codex and Claude Code, while Anthropic explains how to adjust effort, and therefore spending, according to the need for verification, and how to let work in progress finish cleanly when the five-hour limit is reached.

Extensions are taking shape as marketplaces. Anthropic has opened plugin submissions to its directory and added support for MCP 2.0, while MCP usage across its products has increased 110-fold this year according to @ClaudeDevs; Perplexity has launched a Skills Marketplace for Computer; huggingface_hub installs the same skills for Claude Code, Codex, Cursor, or OpenCode with a single command; and Cohere has given Compass Cloud a dedicated MCP server. Skills and MCP servers are becoming common formats that move from one agent to another.

Agents are also entering research. Claude carried out a nine-loop theoretical physics calculation from a one-sentence instruction and simple follow-ups, while a Chinese Academy of Sciences group obtained the core result with GPT-6’s help on certain constraints. Project Swap measures what happens when agents negotiate for humans and finds that the model matters more than the instructions. LeRobot, meanwhile, has published an end-to-end recipe for controlling a humanoid with a learned policy, using low-cost open hardware.