Agent
Governance
Not more agents — the harness that governs them. Agent governance is the difference between an impressive demo and an engineering system you'd let touch production.
What I mean by it
I learned this the direct way. In 2025 my agentic environment (Turbo Flow) scaled to 60+ agents and 215+ tools — and the bottleneck stopped being capability. It was governance: who checks the work, where does context live, and who is allowed to press merge. Turbo Rig is my answer.
The control loop
- ▸ Builder ≠ reviewer. No agent — and no model family — grades its own homework. The reviewer runs on a different family than the builder, every diff.
- ▸ Fail-closed verdicts. Gate can't run? The verdict is REVISE. Reviewer unavailable? REVISE. Silence never passes as approval.
- ▸ Worktree isolation. Every lane builds in its own git worktree; parallel writers can't corrupt each other or main.
- ▸ A written constitution. One AGENTS.md every agent operates under — rules, boundaries, and the non-negotiables (no force-push, no self-merge, secrets never in code).
- ▸ Git-versioned memory. Cross-session context is a git artifact — diffable, reviewable, revertible.
- ▸ Human merge authority. Agents never merge. People do.
Compressed: 215 tools → 60+ agents → 3 roles → 1 independent gate → human merge.
The evidence
7-day window Sept 14–21, 2026 (5 repos) + 116-task controlled study · published by Adventure Wave Labs
The two numbers I'd want any skeptic to start with: 79% of verdicts are REVISE — the gate rejects most agent output, which is exactly what an honest gate should do — and cross-family review moved pass rates from 71.6% to 89.7% in a controlled 116-task study. Governance isn't overhead; it's the yield.
Where this lives
Common questions
Isn't a 79% rejection rate wasteful?
Inverse it: without the gate, that 79% ships to production and becomes incident review. The REVISE loop is cheaper than the rollback loop — and the $416 total review spend for that measured week is a rounding error against engineering time.
What would you tell a team starting with coding agents?
Write the constitution before you grant the tools: what agents may touch, what they may never do, who reviews, who merges. Then make the gate independent of the builder — different model family if possible. Everything else is tuning.
Does governance slow agentic work down?
It slows the merge down and speeds everything else up: less rework, fewer incidents, durable context across sessions. 141 human-merged PRs in a week is not a throughput problem.