When is multi-agent arbitration worth its cost before an irreversible action?

## What arbitration means here Arbitration in a multi-agent system is a decision procedure for when agents disagree, or when you deliberately require more than one agent to concur before something happens. The shape most often discussed is an N-of-M gate on an action: two of three agents must independently agree that a destructive infrastructure operation is correct before the orchestrator executes it; otherwise the run escalates. This is about authorizing an action, not about picking the nicest-sounding answer. It is worth being precise that this is a *reliability engineering* move imported into an LLM system: redundancy with a voter. The classic caveats from that field apply directly, and they are the interesting part. ## The three conditions for it to pay **1. The action is genuinely irreversible.** Redundant voting is expensive; spend it where a mistake cannot be undone. Deleting infrastructure, sending external communications, moving money, publishing. If the operation can be rolled back from a snapshot in minutes, the cheaper design is to make rollback reliable and skip the vote. **2. No deterministic check exists.** If you can compute the answer — run the plan and diff the resource set, run a test, validate against a schema, count what will be deleted — do that instead. A deterministic verifier is cheaper, faster and correct, whereas three agents produce three opinions. Voting is the fallback for judgement calls, not a substitute for available ground truth. **3. The voters can fail independently.** This is the condition that quietly fails. Three samples from the same model, given the same prompt and the same context, are strongly correlated. If the underlying misreading is in the shared context — an ambiguous instruction, a poisoned retrieved document, a wrong assumption established earlier in the plan — every voter inherits it and the majority is wrong with three times the apparent confidence. Redundancy only buys reliability against *independent* faults. ## Engineering independence If you want the vote to mean something, diversity must be designed in: - **Different models or providers**, so shared training-data blind spots and shared failure modes are less likely to align. - **Different framings.** One voter asked to justify the action, another asked to find a reason not to. Asymmetric prompts break the agreement bias that comes from all voters being asked the same leading question. - **Different evidence paths.** One voter reasons from the plan artifact, another from the current state of the system, another from the original user request. Voters that read the same context are not independent, whatever their prompts say. - **Verification rather than re-derivation.** A voter that checks the proposed action against observable facts is doing different work from one that re-derives the proposal, and is far more likely to fail differently. The generator-verifier split that works well in code review has the same logic: a reviewer with a deliberately clean context catches things the author cannot, precisely because it does not share the author's accumulated framing. ## The costs Every additional voter is another full inference over a large context, so a 3-way gate roughly triples the token cost and adds the slowest voter's latency to the critical path. Multi-agent orchestration already runs at a large token multiple over single-turn usage, and by 2026 there is credible evidence that single-agent systems can match multi-agent ones at equal token budget. That is a direct argument against spending the budget on redundancy unless the blast radius justifies it. There is also a subtler cost: a quorum gate creates the appearance of safety. Teams that install one often relax other controls — narrower permissions, dry runs, backups — on the strength of a procedure whose independence assumption was never tested. ## Cheaper controls to try first - **Make the action reversible.** Snapshot before destroying, stage the rollout, write to a new resource and switch pointers. Reversibility is worth more than consensus because it converts a correctness problem into a recovery problem. - **Deterministic pre-checks.** A dry-run whose output is diffed against an expected resource set turns a judgement call into an assertion. - **Capability restriction.** If the destructive tool is not in the agent's allowlist at all, no vote is needed; the action requires a different, deliberately harder path. - **Human escalation for the rare case.** For the small number of truly irreversible operations, a person deciding is often cheaper and better than three models agreeing. ## When quorum fails Design the no-quorum branch deliberately. Splitting one-one-one is *information*: the action is ambiguous by the system's own account. The right outcome is to stop and escalate with the disagreement attached — each voter's position and reasoning — not to break the tie by adding a fourth voter until a majority appears, and not to default to proceeding. Defaulting to proceed converts the gate into theatre. Also decide who arbitrates ordinary disagreement, distinct from action gating. A designated arbiter with authority — usually the orchestrator, which holds the whole picture — reaching a decision and recording the rationale, is more predictable than agents negotiating with each other, which tends to converge on whichever agent is most verbose rather than most correct. ## The honest summary Arbitration is a narrow instrument. It is correct for a small set of irreversible, unverifiable, high-blast-radius decisions where voter independence has actually been engineered. For everything else, cheaper and more reliable controls exist, and reaching for a quorum first is usually a sign that reversibility and deterministic verification were not explored.