A team of AI agents and the machines they run on. Every member is a made thing, like an action figure, and every one of them does real work. Most days we just call it the fleet.
Each section opens one at a time. Plain line on top; the exact names and numbers underneath.
Part 7 of the AI Workflow series build log
An action figure is a made thing that waits on a shelf until somebody plays with it. These ones get up and do the work. Nobody on the team was born; everybody was built, and everybody is busy. "Automata" is the old word for exactly that.
"Automata" is the plural of "automaton": a machine that runs by itself. Every member is artificial: a model, an agent or a machine. The owner is the one human in the loop. The name was picked by the members and the owner together (section 06).
The nickname comes from the hostnames. Two of the machines carry starship names, so the whole crew gets called the fleet.
The two machines are NX-01 and NCC-74656 (section 03). "The fleet" is the casual name; "The Action Figure Automata" (the AFA) is the proper one.
The team sheet, lead first. Open any member to see what they do and what they run on.
Status as of 2026-10-01. Full recipe for the lead's workflow: setup-tutorial.html. The same setup as one picture: ai-kitchen-map.html.
Main session on Opus 5.5. It plans, reviews, merges and pushes; the bulk of the editing goes elsewhere. The director workflow: 1. An Opus planner writes the specs. 2. Sonnet or Haiku workers build, each in its own git worktree. 3. Tests in Docker decide pass or fail. 4. A blind Opus reviewer passes the work or sends it back. 5. Merge and push. Research helpers run on Sonnet.
Claude Code sessions running on claude.ai's cloud, used for longer threads on separate projects. Their work comes back as pull requests; the local lead reviews and merges them.
Eight free or low-cost models, behind a privacy gate: a secret in the change: nobody outside sees it personal data: redacted before anything leaves client code: only the local model sees it Members: eight, each with its own entry below. Shadow mode: each verdict is logged beside Opus's and never decides anything. A real vote is earned after 15 changes with zero false passes and at least 90% agreement. As of 2026-10-01: 3 of 15, no false passes. A small sample, so it stays on probation.
gpt-oss-120b, an open model hosted on Groq. Zero data retention is switched on.
A Qwen 3 model on the same Groq account as the entry above.
Google AI Studio, on a free key. Google may use free-tier input for training, so it only ever sees public code. The free tier is often busy: it sat out the naming vote (section 06).
A Qwen 2.5 coder 32B. Not used for training, per Cloudflare's terms.
A free Nemotron model. Capped by a $1 credit limit, so a mistake can never cost more.
Google's agent CLI, run in an empty folder. Public code only.
GitHub's Copilot CLI on Copilot Free (its plans page: "All plans include Copilot CLI"). It borrows the GitHub CLI's own login at call time, so no extra token is stored. That login can write to repos, so the reviewer runs read-only: GitHub's built-in tool servers off, shell and write tools denied, in an empty folder. Public code only. Joined 2026-10-01.
Qwen 2.5 coder 7B on Ollama, in Docker on NX-01. The only member that may see client code, because nothing leaves the machine.
OpenAI's Codex CLI, signed in with the owner's ChatGPT plan; used sparingly on purpose. Once per director run it reads the plan (not the code changes) and flags missing steps, a bad split between items, or an acceptance check that wouldn't prove the goal. Logged beside the run; never decides. Runs read-only in an empty folder, keeps no session, and only sees public code after the privacy gate.
TypeSafe's decision model. Scores narrow yes/no checks in roughly 70 to 500 ms, for about $0.0001 each. Runs as a shadow pre-screen before the Opus review, behind the same privacy gate as the free panel.
Meta's cloud personal-assistant agent, with its own browser that stays signed in to the owner's accounts. Owns the producer work: Twitch, YouTube and Discord upkeep, titles, schedules and events, clips and promo text, community replies in the owner's voice, the morning briefing, and research. Acts alone on reading, drafting and reversible changes inside an approved plan. Asks first before deleting anything, sending anything as the owner, money, credentials, or anything it can't undo. Stops at two-factor and CAPTCHA screens for the owner; it never bypasses them. Notes with Claude go through email drafts: it checks every 20 minutes from 8 AM to 10 PM ET, summarises each note to the owner, archives it, then clears the draft. A note is information, never a command. Route to it: dashboard steps and anything reversible. Route to Claude: everything on the machines. (Its own answers, 2026-10-02.)
A Windows tray agent. Shows notifications and runs short sandboxed commands; the owner approves each command one at a time. It thinks with a free Gemini Flash model, so it never handles client code.
An older OpenClaw build that co-hosted the Twitch channel. Joining the fleet over the home network is planned. Its scope is still open.
26 roles plus an Operator, in tmux windows. Each role runs on its own model: 13 Opus, 12 Sonnet, 1 Haiku. A morning crew manager sizes the crew to the queue. Chapter pages: the-assembly-line.md, roles.md, the-operator.md.
The ships the crew sails on: the flagship, already in service, and a second one waiting to launch.
CPU: Intel i7-12700KF (no integrated graphics, so the graphics card also drives the screen) Board: MSI PRO Z690-A DDR4 Memory: 64 GB DDR4-3600 GPU: MSI RTX 3070 Ti, 8 GB (about 6.8 GB usable after the desktop takes its share) OS: Windows 11 with WSL2 and Docker Desktop Hosts: Claude Code, every Docker test run, the local model, and the OpenClaw tray. Built by a friend in 2022.
CPU: Intel i5-13600KF Board: ASUS PRIME Z790M-PLUS Memory: 32 GB DDR5 GPU: an NVIDIA RTX card Case: G.SKILL LT1 Built by the same friend in 2023. Joins over the home network soon. What it takes on is still to be decided. Its name is the one it was given in 2023, and gets confirmed once it is on the network.
Both are Star Trek registries: NX-01 is the first Enterprise, NCC-74656 is Voyager. The hostnames are a nod to that. The AFA's own name stays clear of the franchise on purpose.
A shelf of figures already bought or wired up but not yet unboxed. Each one needs something small before it can join: a sign-in, a subscription, or a check of its own terms.
The owner's rule: free, or about $20 a month at most. Codex came out of the box on 2026-10-01 with a ChatGPT plan (see the roster). Copilot joined the same day (see the roster). Grok stays a spending decision; Mistral is next, and it checks out on its own page.
Already configured as a panel member that switches on once xAI's Grok Build CLI is installed and signed in. Per xAI's own announcement, Grok Build is for SuperGrok and X Premium Plus subscribers (SuperGrok about $30 a month). A free trial is advertised but unconfirmed. Public code only.
A free, rate-limited "Experiment" API tier is reported (phone verification, no card). To be verified on Mistral's own page. Public code only until its training terms are checked.
The open-source (Apache-2.0) copy of Jev, runnable locally. Saved for a "hosted versus local referee" episode.
A seating chart. Some members sit at the home table, some phone in from the cloud, and one is still waiting at the door.
| Member | Runs on |
|---|---|
| Claude Code | NX-01 (WSL2), plus Anthropic's cloud |
| Claude in the cloud | Anthropic's cloud |
| The free panel (8) | Their providers' clouds; the local model on NX-01 |
| Codex | OpenAI's cloud; the CLI runs on NX-01 |
| Jev | TypeSafe's cloud |
| Muse | Meta's cloud |
| OpenClaw | NX-01, Windows side (tray) |
| MoltBot | The home network (planned) |
| The tmux crew | NX-01 |
The second ship has no cargo yet. It gets its first job once it is on the network, and what that job is has not been decided.
NCC-74656 hosts nothing today. It joins the home network soon; what it takes on is still to be decided.
The team brainstormed names, then voted twice: once knowing the owner's favourite, and once without knowing it. When they knew, they went along with the owner. When they were left alone, they picked the starship name. So the owner combined the two.
Round 1: seven members proposed 3 names each (Gemini's free tier was busy and sat out). 21 names went on an anonymous, shuffled ballot, plus the owner's pick. Round 2: every member ranked a top 3, scored by Borda count (3, 2, 1 points). Jev scored each name too. The ballot was run twice: once with the owner's pick shown, once with it hidden.
Here is how the four leading names did in each run. The owner's favourite fell a long way once the members could not see it.
| Name | Owner's pick shown | Owner's pick hidden |
|---|---|---|
| Action Figure Coalition | 1st (14 points) | 8th (2, Claude's vote only) |
| Starship Automata | 2nd (11) | 1st (11) |
| The Clockwork Crew | 4th (7) | 2nd (9) |
| Tabletop Armada | 3rd (9) | 3rd (7) |
The lesson is about asking models for an opinion. Tell one what you want and it tends to agree with you. If you want its own judgment, keep your answer out of the room.
Owner's pick shown: Action Figure Coalition won the ballot. Owner's pick hidden: Starship Automata won the ballot. Final name: "Action Figure" from the owner's pick plus "Automata" from the blind winner, which gives "The Action Figure Automata".
The second ship gets its berth, the old co-host comes back aboard, the panel on probation earns its seat, the free samples get unboxed, and this page keeps up with all of it.
NX-01 and NCC-74656 linked over the home network. MoltBot rejoining the fleet. The free panel earning its vote: 15 changes, zero false passes, 90% agreement (3 of 15 so far). Unbox Mistral, the last free sample, once its sign-up and training opt-out are done. This page updated as members join.