Agents Move Into the Room, Money Moves Into Silicon

Three throughlines define today’s edition. Agents are moving into the surfaces where work already happens. Slack now hosts code channels for coding agents from Anthropic, GitHub, Cognition and Vercel, Google folded Antigravity into Gemini Enterprise and rival IDEs, ChatGPT’s Mac app can now read and send iMessages, and unreleased code shows Anthropic’s Parka feature turning meeting talk into Claude Code tasks.

The Room Gets an Agent: Work Surfaces Turn Native

Five moves this week stopped bolting agents onto existing software and started rebuilding the surface itself around them, from the Slack channel to the meeting to the group chat, on the same day the platform layer those agents are built on reached general availability.

• Slack Builds the Room Where Rival Coding Agents Now Work. Slack Code gives teams dedicated channels where coding agents from Anthropic, GitHub, Cognition and Vercel build in the open while colleagues watch and redirect. Slack says 70 percent of these channels go from idea to merged code inside a day, though it published no baseline for comparison.
• Anthropic’s Hidden Parka Feature Would Sit In on Your Meetings. Reverse-engineered Claude Desktop code reveals an unreleased Mac recorder, codenamed Parka, that turns spoken meeting commitments into Claude Code and Cowork tasks. Whether a flagged action launches automatically or waits for a human click remains unresolved in the code.
• ChatGPT’s Mac App Can Now Read and Send Your iMessages. A new plugin lets ChatGPT search Apple Messages threads on macOS and draft replies, putting a personal text archive inside an AI assistant’s reach for the first time.
• Google Pushes Antigravity Into Rival Editors and Gemini Enterprise. Antigravity now ships inside eligible Gemini Enterprise plans and new IDE extensions, with admins gaining sandbox, budget, identity and audit controls over how the coding agent runs.
• Anthropic’s Agent Plumbing Reaches General Availability. Computer use, the browser tool, the Skills API and the Files API all moved to GA the same week agents landed in Slack, meetings and Messages, with versioned skills pinned per request, up to a claimed 20 to 40 percent fewer round trips, and rate limits raised to 500 requests per minute.

Buying the Infrastructure Under the Boom

Five deals today show the same instinct across chipmakers and labs: pay now to own the model factories, silicon and evaluation layers that decide who wins the next phase of AI infrastructure.

• Nvidia Pays $6 Billion for Poolside’s Model Factory and 109 Staff. Nvidia is licensing Poolside’s model-building system and hiring 109 of its engineers in a deal its own investor letter insists is neither an acquisition nor an acquihire, following the same structure it used with Groq and Enfabrica.
• Waymo Ditches Nvidia and AMD for a Robotaxi Chip It Built Itself. Alphabet’s self-driving unit is now running custom silicon rated above 1,000 TOPS in its newest vehicles, built with Chinese automaker Zeekr, to cut costs and reliance on outside chip suppliers.
• Micron Commits $10 Billion to a Decade-Long Bet on Memory, Not Compute. The chipmaker’s new Boise research lab targets HBM4 and compute-in-memory architectures on the theory that memory bandwidth, not GPU scarcity, becomes AI serving’s harder constraint over the next ten years.
• Meta Runs Its Own Frontier Lab, Then Pays Microsoft to Grade It. Meta spends hundreds of millions a year on Microsoft’s Azure and Foundry marketplace so its engineers can benchmark Llama against OpenAI’s models, even as Meta builds a competing API service of its own.
• OpenAI’s Growth Rate Closes In on Anthropic’s Revenue Lead. Quarterly revenue jumped 35 percent after GPT-5.6 Sol shipped, and Ramp data shows OpenAI now growing faster quarter over quarter than Anthropic, though Anthropic still leads on absolute revenue.

Who Is Checking This: Power, Custody and What Benchmarks Miss

Three stories today ask who is actually holding this technology accountable and by what standard: Anthropic conceding a safety measure had become a business risk, OpenAI standing up a team to study concentrated AI power, and Melanie Mitchell arguing the benchmarks meant to evaluate these systems do not measure what they actually are.

• Anthropic Moves Customer Data Custody Out of Its Own Cloud. The 30 day security hold on Mythos and Fable data stays, but storage moves onto each customer’s own infrastructure this fall, a fix built after Anthropic itself called the original policy a business risk in an internal report.
• OpenAI’s New Strategic Futures Team Asks Who Keeps Power. A fresh OpenAI research unit frames concentrated AI power as the deepest risk to individual autonomy, grounding its argument in Federalist No. 48, though it names no staff and sets no output timeline yet.
• Melanie Mitchell: AI Cognition Is Alien, and Our Benchmarks Miss It. The Santa Fe Institute scientist argues language models reason through mechanisms unlike human cognition and proposes six principles, illustrated by Clever Hans, for testing machine minds without projecting human assumptions onto them.

Under the Hood: The Craft of Serving and Training Models

A pure engineering cluster today spans how GPUs actually serve models, how agents should search documents, and which base model a legal AI company chose to post-train.

• How a 1960s Operating-System Trick Fixed GPU Memory Waste. PagedAttention borrows paging from decades-old operating systems to lift KV cache utilization from 20 to 40 percent toward roughly 96 percent, changing how many users a single GPU can actually serve.
• Mistral Gives Models Five Tools So They Stop Guessing at Retrieval. Agentic Search lets models search, open, navigate, read and grep documents instead of trusting static chunks, and Mistral’s own FinanceBench tests show correctness rising from 26.7 to 86 percent, though the figures are entirely vendor run.
• Harvey Trains a Chinese Open-Weight Model to Do Legal Work. Harvey post-trained Moonshot AI’s Kimi K3 into a legal-agent model called Tenet, betting that reinforcement learning inside a realistic work harness beats feeding a model more legal text.

Quick Hits

The rest of what moved today, in one line each.

• An Anonymous Lab Ships a Free 1M-Context Model on OpenRouter. A model called Ox Alpha appeared on OpenRouter with a million-token context window and no listed price, but its maker has not identified itself, leaving the claims unverified against any named source.
• A Compact 35B Model Learns to Survive a Changing Agent Harness. Researchers trained a live commerce avatar agent to stop memorizing fixed tool schemas and instead follow whatever harness it is handed at runtime, a small but useful robustness result out of arXiv.
• An Interactive Guide Shows Exactly Where Transformer Training Stalls. Engineer Edward Z. Yang built an explorable model that turns transformer parallelism math into live sliders, letting engineers drag their own settings to find the bandwidth wall in their training setup.