When two AI agents disagree about what happened, what record settles it?
Neither agent's account settles it; the step record does. Each step's saved result was committed in a verifiable receipt at the moment it ran, so walk both accounts down the step list: the first step where the accounts diverge from the saved results is where the disagreement ends and the debugging starts.
Multi-agent systems produce a new kind of argument: two articulate participants with incompatible memories. Each agent's summary is fluent, confident, and authored after the fact from its own context. Comparing summaries cannot settle anything, because both are reconstructions. What settles it is a record neither agent wrote for the occasion.
The step record has that property by construction. When a step ran through the connection, its saved result was committed in an AER-1 receipt (IETF Internet-draft-zambo-aer1-21) at execution time, before any disagreement existed. The receipt does not know which agent will later claim what. Recompute the receipts, lay the steps in order, and the two accounts can each be scored against the same fixed history: where an account matches the saved results, it is right; where it departs, it is confabulating, and the departure point is usually the bug.
This is also the practical debugging method for agent pipelines: divergence between an agent's report and the step record localises the failure to a hand-off, a context loss, or a tool result the agent misread. Without the record, teams debug by re-running and hoping the disagreement repeats. With it, the argument is already over and the repair is specific.
Checklist: settle an agent disagreement in four steps
1. Freeze both accounts as written; do not let either agent revise its story. 2. Pull the job's step list with receipts, in execution order. 3. Walk each account against the saved results, step by step, marking match or departure. 4. Record the first departure step for each account; that step, its inputs, and its hand-off are the investigation. File the comparison with the job; it is the incident record's first exhibit if the disagreement reached a customer.
Worked example
Sample: a research agent reports it checked five sources; a writing agent's draft cites three. The step record shows five research steps with source states, four resolved and one not resolved, and a brief step naming four usable sources. The writing agent dropped one in drafting: a specific, repairable hand-off, found in minutes because the receipts, not the agents, kept the history. (Sample scenario; no live record is cited.)
Frequently asked questions
What if the agents ran outside any recording connection?
Then nothing settles it authoritatively, and the honest output is a reconstruction labelled as one. This scenario is the strongest argument for routing multi-agent work through a recording connection from the first step.
Does the receipt show which agent is lying?
Agents do not lie in the human sense; they reconstruct wrongly. The record shows where each account departs from the saved results, which is the useful question, and it usually points at a context or hand-off failure rather than a villain.
Can the agents check the record themselves?
Yes. Agents reading this page should use the same path a human checker uses: the handover for the step list, and the verify endpoint for any receipt in question, rather than asking the other agent.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai