Diop Daily #139 — October 2026

When Two Agents Disagree, Who Owns the Replay?

Two agents can complete a transaction while their operators retain incompatible accounts of what happened. One system records an accepted price, another records a changed condition, and both logs appear internally coherent. The dispute begins at the boundary between institutions, where a private log becomes a public claim about authority, action, delivery, and responsibility.

Agentic systems therefore need an evidence object designed for disagreement. It must let counterparties reconstruct the relevant sequence, test which permissions were active, identify the policy that governed each decision, and separate an original event from a later alteration. The record should preserve enough detail for a fair replay while allowing sensitive telemetry to remain under the control of the institution that produced it.

A log tells its owner what happened. A replay gives a counterparty grounds to agree.

The transaction has two memories

A local operator needs rich telemetry: prompts, model outputs, tool calls, retries, policy evaluations, human interventions, and the state of external systems. A counterparty needs a smaller but independently testable account: which principal delegated authority, which action was proposed, what was accepted, what result was delivered, and which evidence supports each step.

These memories serve different purposes. The first supports debugging and security investigation. The second supports settlement, audit, procurement, insurance, and dispute resolution. Collapsing them into one universal log creates a privacy problem. Treating the second as a summary written by the first operator creates a trust problem.

A dispute-grade record should carry at least:

  • Principal and mandate: the person, organization, or service that authorized the agent, including scope, limits, expiry, and exceptions.
  • Configuration: the agent identity, model or policy version, tool definitions, and relevant runtime state.
  • Commitments: the offer, request, approval, or service level that each party accepted before execution.
  • Actions and results: the material calls, outputs, external effects, delivery state, and refusal or escalation decisions.
  • Integrity evidence: timestamps, signatures, hashes, version links, and proofs that reveal alteration or omission.
  • Correction path: the authority that can amend, revoke, or annotate the record and the reason for doing so.

The purpose is to preserve the facts on which a counterparty’s rights and obligations depend without exposing every internal thought of a model. A replay should answer the commercial question without pretending that private reasoning is the same thing as accountable evidence.

Replay is a boundary protocol

Mysten Labs’ October 6 announcement of a Verifiable Agent Arbiter gives this problem a concrete market form. The company describes an evidence layer being co-developed with Google Cloud to help enterprises prove that agents acted within granted authority. Its proposed architecture keeps detailed telemetry in customer-controlled cloud storage while anchoring cryptographic proofs of the record to external infrastructure. The announcement describes dispute replay across companies, incident reconstruction, durable retention, and agentic-commerce evidence as intended capabilities. It is a company announcement about a system in development, not proof that these functions already operate at market scale.

The architectural separation is the important signal. Privacy and independent verification do not require the same storage location. A company can retain confidential prompts and customer data while giving a counterparty a proof that a defined record existed, was not silently changed, and corresponds to the authority and result claimed. The proof does not make the underlying account true by magic. It makes the account inspectable, challengeable, and connected to a method of correction.

NIST’s work on software and AI-agent identity and authorization places the missing prerequisites in view. Its September 29 update reports more than 600 public comments on a concept paper and selects software development as the first implementation use case for demonstrating how agents can be identified, authenticated, and authorized. A replay without identity only shows that an event was recorded. A replay bound to a principal and policy can show whether the event was authorized.

Agreement requires a common fixture

Independent replay does not mean that every party runs the same model again and hopes for the same output. Agent behavior is affected by changing models, tools, external data, permissions, and time. The replay must therefore carry a fixture: the versioned inputs, policy state, authority, relevant external responses, and decision rules that define what another party is being asked to verify.

The W3C Agent Conformance and Benchmarking Community Group describes a related discipline for published agent results. Its proposed work names observable evidence, adversarial fixtures, versioned corpora, engine and corpus identity, dates, and independent reproduction as parts of a conformance record. A transaction replay applies the same logic to a live relationship: the evidence must state the instrument, the conditions, and the boundary of the claim.

The replay may reach several legitimate conclusions:

  • the agent acted within its mandate and the counterparty’s record is incomplete;
  • the parties accepted different versions of a material condition;
  • the action exceeded authority even though the technical call succeeded;
  • the evidence is insufficient, so settlement must pause or move to human review;
  • the record was altered after the event and can no longer support the original claim.

That last category matters. A trustworthy system must be able to preserve uncertainty instead of forcing a binary verdict from incomplete evidence. Institutions can then decide whether to correct a policy, compensate a counterparty, repeat an action, or refuse future delegation.

African institutions need evidence they can carry

Cross-border African commerce and public administration already operate across different legal systems, languages, payment rails, connectivity conditions, and levels of formal documentation. Agents will multiply those boundaries. A dispute mechanism owned entirely by an external platform could make local institutions dependent on a foreign operator’s interpretation of the record.

Institutional sovereignty requires a portable evidence vocabulary. A cooperative, publisher, university, ministry, or payment service should be able to state who authorized an action, preserve the relevant local-language terms, identify the jurisdiction, and present the evidence to a regional partner without surrendering the entire underlying archive. The record should travel with its authority and correction path.

This is also a question of historical method. A source becomes stronger when its provenance, context, and limits remain visible. An agent transaction needs the same discipline. The institution should own the categories by which it describes consent, delivery, refusal, harm, and repair, while exposing enough mapping for a counterpart to verify the claim across a shared protocol.

Where the investable surface is widening

The commercial object is the dispute-grade evidence rail: infrastructure that binds agent identity, authority, action, result, and correction into a record that independent counterparties can test without receiving every private datum.

  • Evidence envelopes: portable records that package mandate, policy, commitments, material actions, outcomes, and proofs with explicit disclosure boundaries.
  • Replay services: controlled environments that reconstruct a disputed workflow from named fixtures and report where the parties agree, diverge, or lack evidence.
  • Authority registries: systems that bind agents to principals, scopes, expiry rules, and revocation events across organizational boundaries.
  • Neutral evidence anchors: services that preserve integrity proofs outside the operator’s private telemetry store without turning sensitive business data into public data.
  • Correction and settlement rails: workflows that connect a replay finding to a refund, amendment, compensation, policy change, or renewed authorization.

Underwriting can begin with operational measures: the percentage of material actions with a named principal and policy; the time required to reproduce a dispute; the rate at which counterparties agree on the reconstructed sequence; the proportion of records that can be disclosed selectively; the number of corrections that preserve the original evidence; and the share of provider or model changes that leave replay semantics intact.

The strongest operator in this market will not merely sell a tamper-evident log. It will maintain the fixtures, identity bindings, disclosure rules, and correction procedures that make the record useful when money, rights, or institutional reputation are at stake. The revenue surface may sit in payments, procurement, insurance, compliance, security, or cross-company workflow infrastructure. The durable asset is the shared capacity to establish what happened.

Run the disagreement before the market does

An institution can test this layer with one workflow that already produces expensive ambiguity: a supplier transaction, a delegated approval, a cross-border delivery, a service-level commitment, or a regulated decision. Name the principal. Freeze the mandate and policy version. Capture the commitment before execution. Record the material actions and external results. Store sensitive telemetry separately from the evidence envelope.

Then create disagreement deliberately. Change the price after approval, revoke the agent midway, return an incomplete delivery, alter a tool response, introduce a language variant, or let one party retain only a partial record. Ask both sides to replay the event. Measure whether they can locate the divergence, identify the authority that governed it, preserve uncertainty, and execute a correction without erasing the original account.

Agents will increasingly act between institutions that do not share a database, model, policy engine, or jurisdiction. Their trust will depend on a record neither party can quietly redefine. The replay is where autonomous action becomes answerable to another institution, and that answerability is the infrastructure on which machine-mediated markets can stand.

Sources