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.

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.

The weekend's agent stories are about what agents leave behind, with Apiiro counting 1,122 public Claude Code permission files that hold secrets, LoreKit arguing that memory files need a shared format, and MaxxBench showing several models' published coding scores fall when task requirements change.

Agent work is moving out of the chat window and into the background, and LangChain, GitHub and Cloudflare each spent November 5 and 6 on the supervision, scheduling and identity pieces that background work needs, with LangChain's notifications in beta and Cloudflare's identity scheme mostly on paper.

Memory stopped being a storage problem this week, with Anthropic publishing an offline pass that rewrites an agent's notes against its own transcripts while a reverse-engineering write-up showed macOS keeping an undocumented, unreviewed knowledge graph of its owner.

Saying no to an agent is moving into defaults and dashboards, with Claude Code 2.1.319 refusing calls its server cannot review and Cloudflare handing site owners a per-agent policy that can end in Block.

Agents now run for hours on someone's behalf, and November 5 showed both sides of the identity question, with Claude Code writing a user's email into its own web requests while Cloudflare proposed a signed way for agents to say who sent them.

The work around AI output is now checking it, with OpenAI needing a paired outcome test to measure GPT-6 Astra's silent sandbagging that its monitors missed and Anthropic releasing a research prototype that proves which model ran.

Meta absorbed Conductor and OpenAI agreed to buy Pierre on the same day, and Conductor's two-week cloud shutdown shows the platform risk that comes with the coding-agent workbench.

An approval rule only binds an agent if it runs somewhere the agent cannot reach, and Pinky shows the design by enforcing phone approvals…

Llama-modchip shows a 284B mixture-of-experts model can run on a 64 GB desktop with no GPU, but its speed comes from a custom file format…

As coding agents move into the chat apps where non-engineers work, the question that decides whether it goes well is who reviews what the…

An agent's saved approval is a file that travels with the repo, so Claude Code's word-for-word allow rules put 1,122 public copies of…

IKnow's real lesson is its split between a loopback-only read-only viewer and an agent manual that must ask before reading, because a…

Google/skills works best as a catalog you pick from rather than a package you install whole, because each skill tells your agent how to act…

Once an agent runs its subagents in the background, the design choice that matters most is what a mid-task correction does to work already…

An agent's notes get worse on their own, so the useful work in agent memory is a scheduled pass that checks notes against transcripts and…