AI Orchestrators

ai , tools , how

An orchestrator is the stuff around a coding agent. The agent writes the code. The orchestrator picks which agent gets a job, gives it somewhere to work, keeps a record of the job outside the session, watches it, checks the result before it lands, and gives me one place to see it all. I didn’t design mine. I added it one block at a time, each block after something broke.

AI-augmented, roughly 70% AI (what this means)

I love AI, and I use it on posts like this one partly to get better at using it. The percent is a rough guess: after rounds of prompting, rewriting and editing, it's very hard to say which words came from whom.

For a wider view, try landscape or desktop.

HUMAN UX CONTROL EXECUTION LANDING GROUND Rich review: annotate artifacts, decide Rich review pages you annotate Overview: what needs you, what is running, what landed Overview the whole fleet, one board Chat: the one agent you talk to Chat talks to you, hands out work Work ledger: durable tasks outside the agent Work ledger Dispatcher: picks agent, model, effort Dispatcher Quota meter: what is left per subscription Quota meter Supervisor: watches from outside every agent Supervisor Worker: a fresh agent with one job Worker 1 Isolation: own git worktree and branch own worktree Runtime: own terminal session own terminal Worker: a fresh agent with one job Worker 2 Isolation: own git worktree and branch own worktree Runtime: own terminal session own terminal Worker: a fresh agent with one job Worker × N Isolation: own git worktree and branch own worktree Runtime: own terminal session own terminal Quality gate: review, test, lint, CI Quality gate review · test · lint · CI Merge path: PR or local merge Merge path PR or local merge Your repos Your repos

Who fills each block

Block Gas City firstmate Claude (native) Mine
Chat: front door You cannot hold ten conversations. One agent holds the whole picture and you talk only to it. Mayor first mate the main Claude Code session Larry, my main Claude session
Rich review A plan, a diagram or a diff cannot be reviewed by scrolling chat. You need a page you can point at. N/A lavish-axi Artifacts, plan mode lavish-axi on trial, markdown files
Overview With a fleet running, no chat and no single page shows what needs you next. You need one board. don't know N/A, you ask the first mate N/A Cockpit
Work ledger Context windows die. Work that lives only in a chat is lost on the next compaction or crash. beads, molecules, convoys tasks-axi backlog session task list only, not durable beads (bd)
Workers One agent is serial: slow, blocks you, confused when tasks stack. A fresh agent per task is parallel and starts clean. polecats, crew crewmates, scouts, secondmates subagents, agent teams Claude subagents, Codex and Grok sessions
Isolation Two agents in one checkout overwrite each other and commit to the wrong branch. worktree per polecat treehouse, pooled worktrees worktree isolation flag treehouse switching from plain git worktree
Runtime You need to see a worker, type into it, and have it outlive whoever spawned it. tmux tmux or Herdr tabs background tasks, in-process Herdr tabs
Dispatcher Subscriptions run out at different times and models differ in strength. Someone has to route. don't know yet crew-dispatch.json rules plus quota-axi floors a model per subagent, Claude only, no quota awareness my judgment plus quota-axi
Supervisor A finished or stuck worker sits silently. The front door cannot watch without blocking you, and its own watchers die with its session. Witness, Deacon patrol agents a bash watcher, outside the agents task notifications, die with the session my watcher daemon, outside the session
Quality gate Agents say done when it is not. Nothing lands on their word alone. Refinery no-mistakes hooks, /code-review no-mistakes on the blog, stronger-model review, I merge

The firstmate column is from its docs, not from running it. The Gas City column is from memory and from my own posts on it. The Claude column is what ships in Claude Code today. Cells that say don't know are cells I have not checked.

Why I hand-roll mine

I run two. Larry runs on the one I built, the “Mine” column in the table above. When I want one off the shelf, I use Gas City.

  1. It fits me: my nudges, my repos, my review habits, my phone.
  2. Building a block is how I learn what it’s for, and that’s how I can judge someone else’s version.
  3. Each block can come from a different vendor, and I pick whichever is best. That matters more when I can’t afford all the expensive tokens I want.

The cost: when a distribution ships something new, I’m behind, sometimes by months. I’m fine with that, because each block is small and cheap to throw away.

Claude will probably absorb most of this

Each block covers something Claude doesn’t do yet, and Claude keeps catching up. Worktree isolation was my own block for a while; now it’s a flag. I’d guess the ledger, dispatcher and supervisor go the same way. No idea when.

When that happens, I delete mine and use theirs.