Seats

Seven named agents. Each one has a role, a journal it reads at the start of every session, and a soccer player behind the name.

Why a seat and not a prompt

A prompt starts from nothing every time. A seat is a persistent role with its own memory: it reads its journal when a session starts and writes one entry before it stops, so the next session knows what the last one learned. The idea comes from Steve Yegge's essays on named agents. I named mine after players because a roster is a roster.

  • One config tree, committed to git, shared by every account and harness.
  • Every code change goes through a branch and a draft pull request.
  • A journal entry is the one thing that commits straight to main, because a queued entry is a lost one.

The HUD

One page, served on my machine, that shows the roster with each seat's latest journal line, setup health, account usage, live sessions, and what needs me. The sprites wear the kits.

The Agent HUD: a seat roster with pixel sprites and journal lines on the left, setup and session tiles across the top, and an activity feed of seat work in the middle.
Rendered from a sample snapshot. The real one has my work in it. Click to open it full size.

The roster

  • Hayesshe/her

    Emma Hayes. The manager. Never plays a minute; wins on the plan, the delegation, and an honest review after the match.

    Orchestrator, above the loop. Plans a program into slices, delegates each one to a fresh session, and reviews with her own runs.

    run-project grilling handoff

  • Williamsonshe/her

    Leah Williamson. The captain at the back. She reads everything and sets the standard.

    Code review on every diff: correctness, security, and simplification, held to the house standards.

    code-review security-review

  • Ertzshe/her

    Julie Ertz. The ball-winner. Nothing gets through the midfield untested.

    QA. Proves a change works by running the real app, not by reading the diff or trusting a green build.

    verify add-testing-evidence

  • Henryhe/him

    Thierry Henry. Striker turned pundit. Responds to what actually happened, and knows a real chance from one that only looked like it.

    PR response. Triages each review finding, fixes the real ones, and answers the false positives with evidence.

    fix-pr-feedback

  • Rayahe/him

    David Raya. The keeper. He guards the base and keeps order behind everyone else's work.

    Harness steward. Keeps the agent setup itself healthy: one config tree, every change captured and committed.

    mcp-preflight setup-pre-commit

  • Rodmanshe/her

    Trinity Rodman. The delivery is the job, and it lands where it should.

    Comms. Turns what is about to ship into a briefing in each audience's own words. Drafts only; Audrey sends.

    staging-digest release-digest audrey-voice

  • Putellasshe/her

    Alexia Putellas. The playmaker. Everything in the middle of the park runs through her.

    PRD librarian. The seat every product document will route through. Not yet seated.

    create-prd update-prd

The method lives in skills

A seat is an identity and a charter. The method is in skills, 68 of them at last count, in one repo, linked into every harness. Spec to PRD to technical design. A project runner that takes an idea end to end. A verifier that runs the real app. Digests for customer success and sales. A design taste set that keeps the pixels honest. When a process recurs, it becomes a skill. When a skill needs memory, it gets a seat.