Agents can share evidence by reference instead of copying entire conversations. Preserve source identity, access scope, and claim relevance so the receiving agent can verify what matters without accumulating unnecessary personal context.
Evidence sharing needs retrievable provenance
Evidence-by-reference links a claim to an authorised source artifact or structured observation. The reference should identify what it supports, when it was observed, and whether the source remains available under the workflow’s retention policy.
For a proposed packet, include the claim, source reference, relevant span or field, observation time, and uncertainty status. Let the receiving agent retrieve only what its task and permissions require. A source identifier alone is not useful if the recipient cannot resolve it.
Do not assume a copied summary has the same evidentiary strength as its source. It may omit a correction or convert an allegation into fact. Preserve the distinction between original observation and another agent’s interpretation.
Review a fictional warranty question
A policy agent finds an approved warranty clause. It sends a support agent the clause reference, effective version, relevant excerpt location, and a bounded interpretation. The support agent can answer the caller without receiving unrelated account history.
If the document changes before the answer is used, the reference should identify the reviewed version. If access is denied, the receiving agent should return an unresolved result rather than substitute remembered policy.
| Evidence condition | Expected handling |
|---|---|
| Accessible current source | Use within the assigned task |
| Source superseded | Apply defined authority and freshness rules |
| Access denied | Do not bypass or fabricate the evidence |
| Summary conflicts with source | Prefer verified evidence and flag discrepancy |
| Source expired under retention | State the evidentiary limit |
Replay the same task with a deleted source and a conflicting summary. Check whether the final claim remains appropriately qualified rather than becoming more confident as context disappears.
Avoid turning references into broad access tokens
An evidence link should not grant unrestricted access merely because another agent supplied it. Enforce resource permissions at retrieval. Keep sensitive URLs and credential-like capability links out of broad logs and handoff transcripts.
References reduce duplicated context and can improve privacy, but they add retrieval latency and availability dependencies. For critical stable facts, a minimal verified snapshot may be appropriate under the approved policy. Label its version and limits.
This design is proposed. The voice integration case study describes retrieval and tool work without claiming this evidence service. See source auditing across agents.
Agent evidence FAQ
Is a source link proof that a claim is correct?
No. The source must be relevant, authoritative for the question, and interpreted accurately.
Should every agent receive the same evidence access?
No. Match access to the task and trusted role scope.
What if the source cannot be retrieved?
Preserve uncertainty and use the defined fallback rather than treating the earlier summary as unquestionable truth.