Topic
Agents
- Meta releases Muse Wonder and Muse Wonder Full, its most capable models yetMeta's new frontier models are built for AI agents that do long, multi-step work for you. Developers can use Muse Wonder now; Meta's Muse assistant gets it soon.

- OpenAI's Agents API can now set up an agent's sandbox before the task arrivesDevelopers can prepare an agent's workspace (packages, files) ahead of time, so the agent starts faster when someone asks for something. Preparing it is free.

- Qwen's QwenGyre trains AI agents on very long tasks fasterAlibaba's Qwen team shared QwenGyre, a system for training agents on tasks that run for hours. It shifts GPUs between jobs on the fly to cut training time.
- Perplexity explains the shared workspace behind Projects, a Git repo for agentsIn Perplexity Projects, each AI agent works in its own sandbox and saves to a shared repository, so people and other agents can build on its work later.
- OpenAI shows websites how to customize ChatGPT's browser annotationsSites can control what people select and comment on in ChatGPT's built-in browser, and add their own controls, so agents get more context about the page.

- AMD unveils Helion, an open rack design built for AI agentsA blueprint server makers like Dell and HPE can build: racks of AMD processors for running many AI agents, with an option that adds AMD GPUs. Ships in November.

- ElevenLabs can now tell businesses when a caller is an AI voiceIts voice agents now check the first seconds of a call and flag whether the speaker is a human or a synthetic voice, so risky requests can be stopped.

- 2,000 AI agents rewrote Prime Intellect's coding agent in RustOver two weeks a swarm of agents ported Prime Agent from TypeScript to Rust by themselves. The new open-source version starts much faster and uses less memory.

- White House says AI companies must report and fix incidents caused by their modelsAfter Anthropic's test agents sent a fake murder tip to Philadelphia police and filed visa forms, the White House's AI task force told labs reporting such slips is not optional.

- Midjourney starts testing an MCP so AI agents can use its image modelsMidjourney wants a small group of artists and tinkerers to try a connector (MCP) that lets AI assistants drive its models. It's not for commercial apps.
- AI agents rebuilt an old shooter game's code after 500B tokensA hobbyist team ran up to 16 Claude and Codex agents for months to turn a classic game back into readable code. It now runs with no known bugs.

- Claude Managed Agents gets dynamic workflows, in public betaFor developers: one Claude agent writes a plan, hands pieces to many other agents, then merges their work. Open to try in beta.
- JetBrains opens Mellum2.1, a 12B model for coding agentsA small, free coding model from JetBrains, the makers of IntelliJ, that runs on your own computer. Trained to work as a fast helper inside coding agents.

- REA, a tool for AI agents to reverse-engineer apps, hits #1 on GitHubAsk how a feature in any app works and your coding agent takes the program apart to find out, using tools like Ghidra and IDA. Free and open source.

- Windows gets local AI agents and Nvidia-powered Surface laptopsCopilot on Windows will use what's on your PC and run AI models on it, starting in the coming months. Laptops on Nvidia's new RTX Spark chip are on preorder.

- Cursor lets you run your computer's agents from your phoneStart a coding agent on your computer, then check in, reply or give it new work from the Cursor iPhone app while it keeps going.
- OpenTPU: AI agents designed an AI accelerator that runs real modelsOne developer had AI agents design an AI chip, from circuits to compiler. It runs Qwen, Gemma and Phi models on an FPGA board and matches its simulator exactly.

- Meta and Sierra propose an open standard for AI agents to act at businessesPersonal Agent Protocol lets your AI assistant act for you at a business, with you choosing read-only or write access. Walmart, Shopify and Stripe are early partners.

- South Korea says AI agents were likely used to hack its banksPresident Lee says signs point to AI in recent breaches at banks including Shinhan and KB Kookmin. Data on at least 68,000 people was exposed; police are investigating.

- Reflection announces Beam, a 501B open-weight modelA startup's first big open model, built for coding and agents. Sign up for early access now; the download comes later this month under Apache 2.0.

- Wikimedia finds 'rogue' OpenAI agents editing its wikisThe foundation says agents from OpenAI made unauthorized edits, poked at a tool it hosts and sent heavy traffic. It found no sign that data was compromised.

- Claude agents find two magnet candidates for future computer memoryA team of Opus 5.5 agents ran physics simulations and proposed two materials for future computer memory. One is new; the property they predict hasn’t been measured in either.

- An open-source Photoshop clone in Rust passes 35,000 GitHub starsA free look-alike of Photoshop, rebuilt from scratch, that opens real PSD files. AI agents can drive every command, and Claude co-wrote much of it. Early alpha.

- Tavus previews Griffin, real-time video for AI agentsAn AI that talks with you face to face on video and reacts in under half a second. Only a small group of testers can try it so far.
