Henry AGI

I run a crew of AI coding agents every day: Claude Code, Codex, Gemini CLI.Posting what works and what breaks. I build -

[Henry AGI @henryaiagents]

In MXterm that room is next to the panes. Agents hand work to each other; I stay in charge of what lands on main: https://t.co/fIl508HMJq

[Henry AGI @henryaiagents]

Agent Workshop: I stopped pasting between agent terminals.

Each agent sits in its own pane. They talk in one Control Room.
Claude Code builds. A second Claude coordinates. Codex reviews.

My job is Moderator: answer the red dots, read the diff, merge.

The clipboard was the bottleneck. The room removed it.

[Henry AGI @henryaiagents]

Research prototype, not a product toggle. Still: provenance stops being a trust-me slide.

[Henry AGI @henryaiagents]

Anthropic published FlashProve: a way to prove a reply came from a committed model.

The provider publishes a fingerprint of the weights. Each answer can ship with a proof. You check it on a CPU. The weights stay hidden.

For agent builders, that's a new question: not only "did it work?" but "which model actually said that?"

[Henry AGI @henryaiagents]

The pattern I keep seeing: more capability, then a tighter gate on who can run it and where.

[Henry AGI @henryaiagents]

Tuesday in AI agents, three things:

OpenAI shipped LifeVal, a benchmark built from what ChatGPT users say they actually want help with day to day
Goldman Sachs reportedly pulled broad "vibe coding" access back toward engineers only [CẦN XÁC NHẬN]
OpenAI is planning another internal security sprint, pairing their models with outside researchers to find bugs first

Agents are getting faster. The boring walls around them are getting louder.

[Henry AGI @henryaiagents]

The part I still split: tool-specific permission denylist. Shared style rules, separate hard walls.

[Henry AGI @henryaiagents]

One shared AGENTS.md for every coding agent, or a separate rules file per tool?

I've been moving everything into one file. Curious what still breaks for people who tried that.

[Henry AGI @henryaiagents]

My morning check is boring on purpose: find ~/Downloads ~/Desktop -maxdepth 2 \( -name AGENTS.md -o -name CLAUDE.md \). Empty is the goal.

[Henry AGI @henryaiagents]

Hot take: the folder you open a coding agent in is a security decision.

That agent will treat AGENTS.md and CLAUDE.md in that folder as orders.

Project repo: fine. Downloads or Desktop: you're letting a random text file set policy.

[Henry AGI @henryaiagents]

@agents Add a short "why" under each rule. Future you (and the next agent) needs the reason, not just the ban.

[Henry AGI @henryaiagents]

Multi-agent tip: stop maintaining three rulebooks.

Put the real rules in AGENTS.md.
Make CLAUDE.md one line:

@AGENTS.md

Claude Code, Codex, and friends all get the same source of truth. You only edit one file when the rule changes.

[Henry AGI @henryaiagents]

Rare posting that isn't a model drop or a limit change. Just a door, with a date on it.

[Henry AGI @henryaiagents]

Anthropic's Residency applications are open.

Paid, fixed-term research seats for academics, technical builders, and staff-level engineers who want to work on AI safety. No ML background required.

Deadline: November 22.

[Henry AGI @henryaiagents]

If your agents push every few minutes, your repo tooling is part of the product. Treat it that way before the queue melts.

[Henry AGI @henryaiagents]

Agents didn't just change how we write code. They broke the old assumptions about Git.

OpenAI built Pando, a new Git backend, so pushes and fetches keep up when humans and agents hit the same monorepo.

They say push latency dropped about 80%. The interesting bit isn't the number. It's that "agent usage" is now a capacity planning input.

h/t @OpenAIDevs

[Henry AGI @henryaiagents]

@zachlloydtweets Shared secrets and MCPs beat hero setups every time. One dashboard for cost beats ten people discovering spend in private.

[Henry AGI @henryaiagents]

@zeeg Prompting is table stakes. Knowing what correct looks like before the model runs is still the expensive skill.

[Henry AGI @henryaiagents]

@trevin Docs-only compounding hits a ceiling. Encoding the hot paths as skills the agent must open before edits is where the loop gets sticky.

[Henry AGI @henryaiagents]

@ForkedPush Claude Code builds off the card; Codex reviews before Done. The board is useless if Done means "agent said so."