One agent's self-improvement is biased
When a single agent both proposes and verifies, judgment easily tilts in its own favor. Just as human organizations separate proposal from approval, splitting self-improvement roles so they check each other reduces bias. The purpose of a multi-agent structure is not speed but checks and balances.
The key is to separate information and incentives so each role judges independently.
Three roles and mutual checks
The proposer makes improvement drafts, the verifier tries to refute them, and the approver decides on deployment. The verifier must aim to break proposals, not pass them, for bias to surface. If the three roles are the same actor or see only the same evidence, the checks collapse.
Full guide: from planning to operations
In planning, define each role's judgment criteria and consensus rules as numbers. For example, require a majority of verifiers to fail at refuting to pass, let a single safety objection act as a veto, and set an approver response SLA. Even with split roles, if all see only the same data they make the same mistakes, so give the verifier a rotating eval set and counter-cases the proposer never saw.
A dangerous failure pattern is collusion. If the proposer and verifier are the same model or the same prompt family, they share weaknesses and the check becomes a formality. To prevent this, compose the verifier from a different perspective than the proposer and assign each verifier a different refutation angle. For recovery, if a majority of verifiers show signs of collusion, discard that round and swap the verifier composition. Handle safety objections as a veto, not majority vote, so no single safety warning is ignored.
On the operations checklist, record each role's judgment separately. Log who proposed, refuted, and approved what, the objection reasons, and the final consensus in standard fields. Use verifier refutation success rate, veto activation frequency, inter-role agreement rate, and cost per round as observability fields. Keep masking rules so data exchanged between roles does not expose personal information.
The continuous improvement loop analyzes proposals that passed but caused problems and refutation angles verifiers missed weekly. Add frequently missed failure types as new verifier perspectives to widen the check network. Multi-agent orchestration should be a living structure that keeps adding verification perspectives to reflect missed cases, not a fixed role layout.
Key takeaways
In short, the value of multi-agent self-improvement is in checks and balances. Separate proposer, verifier, and approver; give the verifier a refutation mission and independent data; and handle safety objections as a veto. Detect collusion and keep adding verifier perspectives to structurally reduce self-improvement bias.