Companies pull back on AI rollouts; Claude's code reviewer ships; Meta drops an 816B open model

Claude Code Review — Anthropic's automated pull-request reviewer is out of preview for Team and Enterprise: install its GitHub App, pick repositories, and new pull requests get read and commented on before a human looks. Review is the bottleneck clients hit once agents write most of the code, so expect to be asked about this one by name. Reviews now land in six to ten minutes instead of about twenty, with at least $10,000 in credits for eligible enterprises until 15 January 2027 — though teams keeping code in the EU under GitHub data residency are excluded from the offer. Anthropic

Vercel Daemons and Drives — Daemons are containers that stay running and pull jobs off a queue; Drives, now generally available, is storage that keeps an agent's files when the sandbox running its code is thrown away. Agent work that spans sessions needs somewhere durable to live, and this arrives on a platform plenty of client teams already deploy their front end to. Drives is on every plan, with 1 TiB of capacity by default, 1 GiB on Hobby, and configurations up to 16 TiB. The New Stack

GrokBot — an always-on agent platform where each bot gets its own chat and its own cloud machine; a Builder.io engineer hired sixteen of them for a week as her staff, one per area of work. It is the most useful public account yet of what always-on agents actually cost and, more importantly, where they silently drop work. She merged 78 pull requests, against a previous best of 35 — then the bots burned $100 over one weekend, mostly waking up and deciding they had nothing to say. Builder.io

AI-DLC — an open set of steering rules from AWS Labs that turns a coding agent into a staged delivery process with human approval gates, running inside Claude Code, Codex CLI, Cursor, opencode and Copilot. When a client asks what structured agentic delivery looks like on paper, this is a free and very concrete thing to react to. Five phases and 33 stages, with 14 agents and 11 workflow profiles; MIT-0 licensed and sitting at 5.3k stars. GitHub (awslabs)

Agent-Native — Builder.io's open-source TypeScript framework for agents that come with an interface, so anything a person can do on the screen the agent calls directly instead of clicking through it. A plausible answer to the chat-box-bolted-on-the-side problem clients keep producing in their first agent project. You define a capability once and it shows up on six surfaces — UI, agent, HTTP API, CLI, MCP and A2A — and every app is itself an MCP server. Builder.io

Muse Spark Open — Meta's new open-weight model aimed at coding and agent work, free to download, with 816 billion parameters of which 42 billion fire per token and a one-million-token context window. The "can we run something good ourselves" question now has a US answer, but read the licence before promising a client anything. Meta says two Apple M5 Ultra Mac Studios will run it on a desk, and any product built on it must show a "Built with Muse Spark Open" credit. Meta AI Research

Nemotron 3.5 Super VL — NVIDIA's open 120-billion-parameter model for long-running agents that need to read screenshots, PDFs, tables and charts as well as text, released alongside two smaller Bolt models for laptops and edge devices. Weights you can post-train on your own data and keep on your own infrastructure is the pitch regulated clients actually respond to. Scale AI post-trained it for EY's M&A due-diligence agent, and the 74B Bolt variant squeezes into 24GB of memory using native 2-bit quantization. NVIDIA

Anthropic — told investors that revenue for July to September came to $16.8bn, up almost half on the previous quarter's $11.5bn, as it prepares to go public. This is the figure behind every "is this vendor durable" question, and it suggests the business is now close to paying for its own compute. The company reported $1.70 of revenue for each $1 spent on computing power, 11% adjusted operating margins, and said distillation by rival labs — 5 to 7% of revenue — was cut back materially in the quarter. Simon Willison

OpenAI gives GPT-4o users half the notice it promised. The GPT-4 generation leaves OpenAI's API on 14 January 2027 — 4o, 4o mini, 4.1, 4.1 mini and anything fine-tuned on them. The notice went out this week, which is three months, where the company's own published policy promises at least six. If a client still runs production traffic on 4o, that migration is now dated work, and the behaviour differences matter more than the model ID. GPT-6 Luna, the suggested replacement, lists at $0.10 per million input tokens — but a prompt over 272,000 tokens doubles that input rate for the whole request. The New Stack

GhostAction — a credential-theft campaign that used around 1,900 hijacked GitHub accounts to plant innocuous-looking workflow files in over 30,000 repositories, scraping CI secrets, cloud keys and API keys for Anthropic, OpenAI and OpenRouter out of both code and git history. Your agent provider keys live in exactly those places, and keys deleted months ago but still valid in history were fair game. The 8 and 9 October push tried to extract 3,174 secrets; by 5 October only 124 of 772 known affected repositories had been cleaned up. The New Stack

FTC — has issued civil investigative demands — a form of subpoena — to OpenAI and Anthropic, seeking internal records on security breaches, safety safeguards, marketing claims and data privacy policies. The questions being asked (who authorised the agent, within what limits, can you produce the record) are the ones your clients' auditors will ask next year. The inquiry began over the summer; the FTC chair also co-leads the administration's Super Intelligence Force, and an AP-NORC poll found 64% of American adults think AI is moving too fast. Techstrong.ai

Universal Cart — a shopping basket that lives inside Google Search and the Gemini app, holding items from several retailers at once and taking payment through Google Pay, with each merchant still the seller of record. Agent-driven buying has reached ordinary consumers, which is how the "will customers still come to our site" conversation gets started. The rollout plan names Canada, Australia and the UK next and says nothing about Europe, where Brussels fined the company €890M under the DMA in July. The Next Web

ChatGPT's Chrome extension — can now look at your open browser tabs, pull details out of them into a sheet or a draft, and on the Work plan control tabs to fill in forms, send messages or download files. A browser agent inside the browser everyone already uses is a quieter route into enterprises than any standalone AI browser was. It follows the discontinuation of OpenAI's Atlas browser and closely mirrors Google's auto browse feature in Gemini. The Verge

Modal — rents the throwaway cloud machines that a great many agent products quietly run their code inside. It says it has reached a billion dollars of annualised revenue and taken a $750M round led by Accel that prices it at $15.75B. Sandboxing has stopped being a feature and become an infrastructure category, which is useful ammunition when you argue a client's agents should not run on someone's laptop. Its Sandboxes product, launched in 2023, is a $300M run-rate business on its own, with close to 3 billion of them started. Latent Space

How Codex Happened — a long account from Every of how Codex went from a cloud agent that failed to find users, through a pivot into the terminal and the code editor, to the product that ended up absorbing ChatGPT. Good background for the client question of why your developers keep switching harnesses: the surfaces moved faster than the models did. Codex passed 5 million weekly active users by June 2026 and merged into the flagship ChatGPT desktop app the following month. Every

“The first real, massive ‘feel-the-AGI’ moment for me was [when] I saw someone running a reasoning model that was connected to a terminal”