Fleet Series 01 — The Big Picture: A Self-Hosted AI Software Factory

Working notes for a planned episode — an outline, not an article. Details will shift when it gets scripted. Series index: Building the Fleet.

The overview episode: the whole fleet on one map. Everything else in the series is a zoom-in on one box of this diagram.

The hook

Everyone demos an agentic coding loop. Almost nobody shows you what it takes to run one unattended, on your own hardware, against real repos, and still trust the output. This is the tour of everything around the loop — which turns out to be nearly everything.

What the viewer walks away with

• The loop itself is ~5 lines; the surrounding system is the actual work
• A mental map of the layers: task store → planner → orchestrator → workers → gates → provenance → forge, with local models and sandboxes underneath
• Why every layer is either an evaluator, a substrate, or a paper trail

The map (one diagram, spend the whole episode walking it)

• Forge (a notebook in Nous, my note-taking app): where task specs, features, and dependencies live — the fleet's shared memory
• Architect: goal → framing proposal → tree of small, testable leaf tasks
• Orchestrator: executes the tree in waves — everything unblocked, verify, reconcile, repeat
• Workers: one leaf each, in an isolated workspace, behind a task-store port (Forge, GitHub, git-bug)
• Ensembles: fan-out to many models, combine with quorum — review, testing, audit
• Gates: fail-closed checks between "a model did a thing" and "we keep it" — suite gate, review findings, sign-off
• Provenance: verdicts and journals written back as git objects in the target repo
• soft-serve-with-sprinkles: the self-hosted forge that stores the repos and renders the provenance
• LLM router: local model fleet behind one OpenAI-compatible endpoint — the fleet's power supply
• gaol: container/VM isolation so parallel execution and untrusted code are routine, not scary

Beats

• Cold open: kick off a real epic, walk away, come back to landed commits with verdicts attached
• The reframe: generator–evaluator, not "autonomous agent" (pull from the lightning talk)
• Walk the map top to bottom, 60–90 seconds per box, each ending with "deep dive in episode NN"
• The dogfood story: the pipeline built large parts of itself (35/35 tasks on its own construction)
• Close: the three jobs — evaluate, isolate, remember — and the series roadmap

Demo ideas

• Time-lapse of one epic flowing end to end: task tree → waves executing → gate verdicts → provenance in the forge UI
• Split screen: task spec on one side, worker commits landing on the other

Source material

• Lightning Talk — The Loop Is the Easy Part — the argument skeleton
• On Agentic AI Coding Loops — the thesis prose
• The internal development-workflows playbook — the prose version of the map

Open questions

• One diagram or a build-up sequence? (Probably build-up: add a box per beat)
• How much homelab hardware detail belongs here vs. in the router episode

Closing zoom-out — the two human loops (now its own episode: 11)

The three-loop framing that used to live here in full — coding loop (minutes, the fleet runs it) ⊂ feedback loop (hours, the human reviews output and updates specs and rules) ⊂ outside loop (days, real users drive change) — grew into its own episode: Fleet Series 11 — The Two Loops You Still Drive. In 01 it stays only as the closing beat: after walking the map top to bottom, zoom out once and draw the two human loops around the whole diagram, teeing up 11's answer to "so what do you do all day?" (the human's job = intent + accountability).