01The Loop, whole first
Everything below serves this one cycle. If you remember one image, remember this one.
Spec → build → cross-model gate → human diff review → merge. Agents never merge: the guard denies agent pushes to protected main, so the merge button is human-only by construction, not by promise.
02The Triangle: three agents, distinct jobs
Model quality converged in 2026; the harness and the loop are the differentiator. So the rig is not "more agents". It's three seats, filled by strength, swappable any time.
Builder lane
The agent that writes the change: spec in, branch out, in its own isolated worktree. Any headless CLI: ZCode, Claude Code, Codex, Grok, Gemini. You bring your own subscriptions; the rig orchestrates them.
Reviewer lane
A different model family reads the diff read-only and must end with a parseable APPROVED or REVISE. Builder ≠ reviewer, the AI equivalent of author ≠ approver. If the reviewer is down, the gate fails closed: no verdict, no merge.
Check lane
Deterministic checks first: tests, type checks, lints. They cost $0 and speak for nobody. In the audited week 24 of 868 gate runs were check-lane verdicts.
What this bought, measured: in the audited week (Sept 14–21, 2026; gate live Sept 16–21) the gate ran 868 reviews and sent 687 back (79.1%); 141 PRs still shipped, every merge held by a human. Every number re-derived from the primary logs.
03The Spine: one repo of contracts
The Triangle's agents share one backbone, so every harness reads the same rules and every machine sees the same memory.
~/rig or wherever you clone it ├── AGENTS.md the constitution: invariants every agent must obey ├── memory/ git-versioned agent memory: one brain, many bodies ├── specs/ work on disk, not in chat scrollback ├── runbooks/ how the humans operate the rig └── scripts/ review.sh (the gate) · wt.sh (worktrees) · …
AGENTS.md, once
One written constitution, read natively by ZCode, Claude Code and Codex (CLAUDE.md → AGENTS.md symlink). Change the law in one place; every harness obeys the new law on its next run.
Memory in git
Cross-session, cross-machine agent memory as plain markdown in git: diffable, pullable, backed up with the code. If a vector index helps later, it consumes these files; it never replaces them.
Plane-portable
The spine is git, so the same rig runs on the laptop, an always-on VPS or a Codespace, staying in sync by git pull. Your laptop stays the cockpit; it stops being the only place the loop can run.
04The invariant: agents never merge
Every rule above bends to your setup except one. The guard denies agent pushes and merges to protected main. That is why "141 PRs merged, every one by a human" is an invariant claim, not a hope: the audit counted the merges; the human-hold is enforced by the guard (and stated honestly in the evidence page's limits, because the audit did not re-derive merger identity per PR).
Fail-closed everywhere
No reviewer available? No merge. Gate can't parse a verdict? No merge. The failure mode of the rig is "stops", never "lets it through".
Isolation by default
Every change builds in its own worktree. Parallel agents never share a checkout, so their writes never collide and a broken lane dies alone.
Everything on the record
Gate verdicts, costs and durations land in a log. That log is what the audited week was re-derived from; the receipts are public.
05Try it today
The full rig is in private beta (request access). The gate itself is available now: rig-lite is the free, open kit from the Turbo Flow lineage: a review gate you can install in minutes and point at your own agents.
GET RIG-LITE ↗Lineage: Turbo Rig is based on Turbo Flow, the open-source rules layer; published by Adventure Wave Labs.