Agent builder
Checker live · builder in progress
Runs on Brain →Watches how AI is used across my projects, flags waste, spots where a new agent or script would help, and writes each one up as a proposal. Nothing changes until I approve it.
- 11,000+model calls read in one scan
- 3waste detectors live
- 88tests on the checker
- 6reviewer verdicts for auto mode
HOW IT WORKS
Architecture
Code gathers the evidence and decides nothing. Claude scores each idea, code checks every citation, and I approve or deny.
Interactive diagram. Select a card to highlight what it feeds and what feeds it; press Escape to clear.
UNDER THE HOOD
What it actually does
Abstains rather than guesses
No evidence, no score. A score without a citable line is marked unknown, never invented.
Citations checked by code
Every line a proposal cites must exist word for word, or that score becomes unknown.
It reported its own blind spot
Its first live scan read 11,000+ calls, found none tagged by task type, and said so instead of guessing.
Nothing hardcoded
Every threshold lives in a rules file. A missing rule is a visible error, not a quiet default.
Not checked is not passed
A rule that cannot be checked writes nothing and never closes a proposal.
A hard write scope
Never push, never delete, never touch a key file. Logs are append-only, checked line by line.
Every agent ships with limits
Its own permissions, a run cap and a kill switch. I start it by hand.
Auto mode needs unanimity
Off by default. Building on its own needs high safety plus six reviewer verdicts that all agree, and unsure counts as no.
Tested against an answer key
Four planted cases had their expected verdicts written down before the scan ran.
HOW I DESIGNED IT TO WORK
It proposes. I decide.
The brief it started from. Every change still waits for my approval unless both reviewing AIs agree and I have switched auto mode on.
“Analyzes projects, determines where an agent could be useful, builds the agent, puts it to work when relevant.”
The original brief
WHAT'S NEXT
Next, in order
Build four recent rulings into the code, starting with closing a proposal only after two clean checks.
Land two pending patches and restart the local-model logger.
Finish the fourth detector, for local-model health, once real failure data exists.