Give a headless brain one directive and it turns it into a mission that a team of AI agents plans, builds, verifies, re-plans when they hit problems, and iterates with the client until it's accepted.
Real recorded mission: 63 tasks (62 done), 56 findings, 18m — replaying at 4× below.
A directive becomes a mission; the brain decomposes it into a dependency-ordered task queue; the agents pull ready tasks and execute them; their structured findings feed a two-tier re-planner; the mission converges when the client accepts.
CLIENT (you, or a modeled product-owner agent) │ directive ↓ ↑ accept / feedback → next sprint ▼ │ LEAD ── research → design → build-core → build → test ∥ secops ∥ perf → integrate → docs → retro (orchestrates, the dev team (one role per phase) re-plans, reworks) SCRUM (standups · stall call-outs · nudges) └── findings → reflex fix+verify ∥ lead supersede / re-architect → converge
Recurring failure signatures (the same finding, again and again) and clusters of similar lessons are swept into skill proposals: an LLM drafts corrective guidance plus a reusable skill, Shep announces the pending proposal at standup, and the operator approves or rejects it — from its own Proposals tab (a live count badge) or corral-admin proposals. Approval promotes the guidance into vetted memory and a versioned skill artifact; every later mission's instructions carry the top vetted lessons (fence-wrapped, clearly labeled, capped at 3) so the herd starts each mission already warned. And the loop watches its own efficacy: if the same signature keeps recurring after promotion, a revision proposal reopens for the human to reconsider.
A repo that runs with corralai can carry its working knowledge as a markdown corpus in the repo itself: CORRAL.md at the root as the entry point, docs/corral/*.md as the corpus. The same corpus serves four readers — developers read it as onboarding docs, any developer's coding agent queries it conversationally, the herd itself searches it before working and extends memory as it learns, and it grows the way code does — through ordinary pull requests, where code review is the trust gate for knowledge exactly as it is for code.
The learning loop closes the circle: skills the herd proposes and a human approves land in the same corpus — herd-discovered knowledge and developer-written knowledge accumulate in one place, under one review gate, readable by humans and queryable by every agent that joins.
Nothing about a finished mission is thrown away: every task's claim and completion, every finding and its resolution, every command an agent actually ran, and the event log itself survive indefinitely. A Completed tab lists past missions — directive, duration, task/finding counts, and (best-effort) what got learned from them — with a detail view per mission and a ▶ replay button. Replay is read-only: it reconstructs the whole build from durable rows and plays it back on the same corral canvas, at up to 16×, with a scrub bar.
This hero IS that player — the exact same code, replaying a real mission.
go test ./... go run ./cmd/corral # MCP /mcp/ · health /healthz · live UI / · on 127.0.0.1:9019
Open http://127.0.0.1:9019/ for the live herd + Progress tab (dev: auth off). To watch the whole loop end-to-end on one command (bundled GPU Ollama):
cd deploy/demo make demo-mission # directive → team builds it → re-plans → client review → converge