AI agent orchestration, governed

Context

• Agent demos are easy and agent operations are not. The gap is not model capability; it is authority, verification and recovery.
• An agent that is usually right is dangerous precisely because it is usually right - review attention decays exactly when it is most needed.

What was actually wrong

• Autonomy without a gate produces confident, plausible, wrong output that reads as finished work.
• Without a durable record of what ran and why, nothing can be audited, rolled back or trusted after the fact.

Approach

1. 01Write the operating charter first: who may decide what, which actions are gated on explicit human approval, and what happens when a gate is hit mid-run.
2. 02Treat every completion claim as false until independently checked. Read the diff, not the summary of the diff; hit the URL rather than accepting that it is live.
3. 03Use adversarial verification for anything load-bearing: independent reviewers whose job is to refute, with a majority needed to let a claim survive.
4. 04Make the record durable and reconstructable - an append-only journal from which the operational state can be rebuilt, so the working store is disposable.
5. 05Keep irreversible actions behind per-instance approval. No batch pre-authorisation, because blast radius is not additive.

Architecture

• Coordinator plus scoped workers with explicit file ownership and isolated working trees
• Append-only journal as the system of record, with replay-based reconstruction
• Structured verification stages between generation and delivery
• Local-first operation with outward-facing actions gated

What this demonstrates

• Designing autonomy with the safety properties decided before the capability is switched on
• Verification engineering: making correctness a property of the process, not of the operator's diligence
• Auditability as a first-class requirement

Related