Blog

Thoughts, notes, and documentation
← All posts

Retiring the Brain: 8,000 lines deleted, the workflow stays

ProjectsAIAgentsToolsWorkflow

Two weeks ago I published a spotlight on the Brain — the live status board my AI agents and I used to run this monorepo. Yesterday I asked whether I still needed it. Last night the agents deleted it: three PRs, about 8,000 lines removed, shipped in one evening. This is the story, because the numbers that killed it are better than the app was.

The Brain viewer — retired July 2026

The Brain was a markdown protocol with a dashboard on top. Every project carried a BRAIN.md — what is this, what's in focus, what changed lately — and a private viewer rendered them all as cards with live GitHub issues mixed in. It worked. More importantly, it trained the whole multi-agent routine: identity cards per project, a hard three-line cap on "current focus", a dated close-out line after every session, and issue-first work in isolated git worktrees.

But when I asked Claude "should I retire this?", it ran a seven-agent audit of the repo and came back with numbers I couldn't unsee:

  • 59% of all commits in the last 90 days touched a BRAIN.md file.
  • Of those, 58% were pure dated log appends — no knowledge, just compliance.
  • Every one of those appends triggered a full production deploy.
  • The dashboard itself? I'd opened it on four days in ten weeks.
  • Roughly 7,800 lines of infrastructure were maintaining 627 lines of content — twelve lines of machinery per line of truth.

I got a second opinion from Grok, which argued for keeping the card files and the habits while retiring the app — and it was right about that. The merged plan: GitHub Issues own all work (that migration had already happened), the per-project card files stay for agents and humans arriving at the code, and everything else goes.

Execution was the fun part — three agents in parallel worktrees:

  • #1307 — card-only pushes stopped triggering deploys, labels became automation-applied (humans never curate them), issue templates went in, and the one external consumer of a Brain endpoint got quietly re-plumbed first.
  • #1308 — every irreplaceable fact buried in a BRAIN.md (deploy invariants, "this data is gone forever" warnings, do-not-reopen decisions) moved into READMEs and a decisions log; a one-page spec and a ~100-line lint replaced the old stack.
  • #1313 — the demolition itself: −8,014 lines.

What survived, because it earned it: the card files (now just "project cards", same filename), the three-line focus cap, the close-out line — written only when something actually shipped — the git union-merge trick that lets parallel agents append to the same file without conflicts, and the little lint that keeps all of it honest.

The lesson I keep relearning with AI agents: the durable asset is never the app. It's the protocol the app forced you to practice. The Brain's job was to teach a multi-agent team how to leave the campsite clean — and it did. Work lives in GitHub Issues now; the cards stay behind for whoever — human or agent — shows up at the code next.