Agent builders are separating the part that decides from the part that acts, with Google's Antigravity Orchestrator passing a plan between Carbon and Hydrogen, IBM's LLMON training Granite to obey only bound instructions, and a new paper finding that weak actors make strong judges, while a PostgreSQL bug shows what happens when data gets read as something it is not.

Agent builders are separating the part that decides from the part that acts, with Google's Antigravity Orchestrator passing a plan between Carbon and Hydrogen, IBM's LLMON training Granite to obey only bound instructions, and a new paper finding that weak actors make strong judges, while a PostgreSQL bug shows what happens when data gets read as something it is not.

Agents are shedding the single vendor account, with Brave selling search to a wallet instead of an API key, Edge0 putting a 145B model on one NVIDIA GPU plus system RAM, and gateway repos routing Claude Code to other providers, on a morning when a Claude sign-in incident showed what one account costs.

The model that runs an agent's loop is sliding down the price ladder, with Google's Hydrogen, Tether's BotPsy and the SteerDistill paper each pushing agent work onto cheaper or smaller models, and Microsoft's own tables showing how little the top tier sometimes buys.

The places agents act are starting to set terms at the door, with United refusing agents that book as the traveler and Australia drafting an AI regulator after an OpenAI agent got into a government data portal.

The builder launches of November 11 give agents a place to check their own work rather than a smarter brain, led by AWS shipping a local copy of AWS for coding agents and Google pricing a yes-or-no judge at a dime per million input tokens.

AI now produces code and bug reports faster than people can review them, and this week's answers ship as context packs and contribution policies rather than smarter models.

Today's AI news is about the parties you never see in the loop, with Notion customers learning their AI chat transcripts sat at a breached inference vendor, GitHub shipping Claude Fable 5.2 with Anthropic retaining prompts by default, and DeepMind putting a proof checker between AI-written verification and Signal's cryptographic code.

This weekend's agent tools make deliberate choices about what an agent holds, with Google's Elephant Goldfish Model keeping reviewers out of the conversation, billion-context folding old turns into summaries, and llama-modchip squeezing a 284B model and its cache into 64 GB of RAM.

Brave's x402 search proxy swaps the API key for a per-query wallet payment, which moves an agent's security boundary from who it is to how…

Vercel-labs/ui makes a control's accessible name the handle an agent uses to operate it, so building a UI agents can drive safely becomes…

OpenFilm treats every shot as a web page with one rule (the same time always draws the same frame), which lets an ordinary coding agent…

LLMON moves the line between instructions and data out of the prompt wrapper and into the format the model was trained to read, which makes…

United's ban targets agents that log in as the traveler, and the builders who keep their access will be the ones who stop borrowing user…

Shadcn/improve makes a plan file, not a model choice, the handoff between an expensive advisor and a cheap executor, and the part worth…

Lot-map turns a coding-agent session into a map you can read at a glance, but because its hooks receive every tool call's full input, you…

Gemini 4 Hydrogen's flat price across a 1M-token context makes a cheap tier the sensible default for agent loops, but only after you run…