Maker-checker — the four-eyes principle from banking and aviation — requires that every consequential action be produced by one party and approved by a different, independent one. Ported to agents, it is the minimal architecture for trustworthy autonomy: the agent that does the work is never the one who signs it off.

The requirement is structural, not advisory, for one reason: a single agent grading its own work has exactly one failure mode — self-assessment bias — and it fires constantly. Models are known to be poor judges of their own errors; confidence and correctness correlate weakly, and a generator asked "did you do this right?" will mostly say yes, convincingly. Separation of duties breaks the correlation by construction: the checker sees only the artifact and its acceptance criteria, never the maker's reasoning, confidence, or conversational lobbying. Information asymmetry is not a bug in the design — it is the design.

The pattern generalizes along a dial of independence. At minimum: a separate verifier agent with a different model or materially different prompt, reading the artifact cold. Stronger: a deterministic checker — tests, linters, schema validation — as the first gate, with the model-verifier second. Strongest: a human gate for the actions you cannot afford to be wrong, with everything else promotion-gated on the automated chain. The right setting is per-action-class, driven by blast radius, and written down.

Implementation Rules

• Isolate the packets: the checker receives acceptance criteria and the artifact — not the maker's narrative. Shared logs recreate the bias you paid to remove.
• Bind approvals to evidence: every PASS cites what was checked, when, on which revision — approvals attach to a state, not to a vibe.
• No self-promotion: the maker cannot modify the checker's criteria, and the checker cannot edit the artifact — only accept, reject, or escalate.
• Audit the rejections: checker disagreements are your highest-signal data — mine them for recurring failure classes, and for calibrating when autonomy can widen.

Maker-checker is also the pattern that makes delegation legible to organizations: auditors understand four-eyes control instantly, which is why agent platforms that adopt it navigate procurement and compliance while platforms that promise "fully autonomous" stall in review. Trust, it turns out, is an architecture.