The operational problem
A workflow can produce a persuasive answer while an external update fails or finishes after a timeout. It can also resume after an interruption with stale evidence. The review question is concrete: which action was permitted, against which object, using which evidence, and what changed? Design the journal around these transitions rather than trying to reconstruct them later from prose.
Evidence to collect before acting
- Request and sources
- Keep a stable workflow ID, permitted task, source references and observation times. Link to protected evidence rather than copying customer records into public logs.
- Approval record
- Retain the authorized reviewer, pending request identifier, exact action scope and decision time. If the action, destination or amount changes, reassess whether the approval still covers it.
- Execution and read-back
- Record the tool and step identifier, attempt time, destination reference and observed result. Keep credentials and unnecessary personal data out of the journal.
A practical workflow
Step 1
Use explicit states
Separate proposed, awaiting approval, rejected, executing, verified, failed and outcome unknown. A rejected or expired request must not silently become authorized.
Step 2
Bind the decision to the action
The trusted execution layer should validate the pending request and arguments before invoking a privileged tool. Microsoft’s approval example binds an approval response to a pending request in the same agent session.
Step 3
Verify the external outcome
After a write, read the destination state or authoritative receipt. If the connection drops, record uncertainty until reconciliation establishes what happened.
Step 4
Recover with the same identity
On restart, load saved steps, recheck evidence freshness and reconcile uncertain writes before retrying. Saving a journal does not by itself prove that crash recovery or duplicate prevention works. Test those behaviors separately.
Worked example · Illustrative
Illustrative example, not a customer result: an operator approves one payment-status update for DEMO-104. The request times out after submission. The journal records “outcome unknown,” not “failed, retry immediately.” A recovery step reads the destination using the saved order and processing reference. If the update is already present, it records the verified result without repeating the action.
Exceptions that need a separate decision
| Situation | Handling approach |
|---|---|
| Approval replay | Reject a decision that does not match the current pending request. A previously approved action should not authorize a different order or amount. |
| Revoked access or stale evidence | Stop the affected step and show which source or permission needs attention. An older successful observation must keep its original timestamp. |
| Partial completion | Preserve successful steps and identify unresolved ones. Avoid labeling a multi-step task complete merely because one tool call succeeded. |
How to measure the result
Test rejected and mismatched approvals, revoked credentials, duplicate deliveries, process termination and timeouts after external writes. Measure verified outcomes, unresolved attempts and recovery time separately. An audit trail supports investigation; it is not evidence of a security certification or a guarantee that a system is safe.
Common questions
Is an agent conversation an audit trail?
It can be part of the evidence, but it often omits enforced permissions and destination state. Link the conversation to structured approval and execution records.
Does every read-only step need approval?
Not necessarily. A narrowly scoped read can follow a standing policy. Use action-specific review where the organization requires it, with authorization enforced by the tool or execution service.
Primary sources
Platform documentation supports the distinctions described here. Workflow controls and measurement recommendations are design guidance from LINKIO.ONE, not evidence of a completed LINKIO deployment.