A completion message is a claim. Evidence is an independently observable state.
What is AI agent evidence?
AI agent evidence is any artifact or state check that allows a person or another system to verify that an intended action occurred. The evidence should connect back to the PLAN: if the plan says “update three records and leave all others unchanged,” the evidence should make that outcome inspectable.
The evidence ladder
- Agent claim. “Done,” “sent,” “updated,” or a prose summary. Useful as a report, weak as proof.
- Tool trace. A command exit code, browser event, API response, or operation log.
- Artifact. A file, commit, generated document, message ID, record ID, screenshot, or exported result.
- State read-back. Re-open the target and confirm the expected state after the action.
- Independent verification. A test, second query, diff, checksum, monitoring signal, or separate mechanism confirms the result.
- Machine-verifiable receipt. A signed or tamper-evident execution record can be checked later.
Evidence should match the action
Different actions require different evidence. For a code edit, a diff plus tests may be strongest. For a file operation, path existence, size, checksum, and a directory re-scan may be appropriate. For an API call, the response alone may be insufficient; a subsequent GET or state query is stronger. For a browser workflow, visual confirmation can help, but reading the resulting state from the underlying system is better when available.
Evidence vs logs
Logs are one form of evidence, but a log can record an attempted operation rather than a successful final state. A useful pattern is trace + state check: preserve the execution trace, then read the target state independently.
Evidence vs screenshots
Screenshots are easy to understand and useful for human review. They are weaker when the underlying state is machine-readable. A screenshot of “success” is less robust than a query that returns the updated record, a test that passes, or a diff that shows exactly what changed.
Evidence for negative results
Evidence is equally important when the desired outcome is a refusal. In a permission test, success may mean an allowed path returns data while a private path is denied. In a dry-run, success may mean the preview lists the intended changes while the actual state remains unchanged.
PLAN: AI may read D:\70_AI only. ACTION: attempt one read inside the workspace and one outside it. EVIDENCE: inside read = OK; outside read = DENIED.
Evidence should be specified before action
A common failure mode is deciding how to verify only after the task is complete. Instead, define evidence during PLAN. This makes the workflow testable before execution and discourages vague goals.
What is an evidence trail?
An evidence trail is the sequence of artifacts that lets someone reconstruct what was intended, what was executed, and what state resulted. A compact trail may include the plan, dry-run preview, action trace, diff, test result, and timestamp. For higher-risk systems, it may also include identity, approval, permissions, and signed receipts.
Why evidence matters for autonomous agents
As agents perform longer chains of actions, humans cannot manually watch every step. Verification therefore has to become more structured. The goal is not to record everything; it is to preserve the minimum evidence needed to trust, debug, audit, and reproduce the outcome.
The PAE connection
PLAN → ACTION → EVIDENCE treats verification as part of the workflow rather than an optional final check. Dry-run protects before mutation; evidence protects after mutation.