Governance / FIELD NOTES
Human approval, agent authorization and execution accountability
Why an approval click, an access decision and an observed mutation need separate evidence—and how to connect them without adding unnecessary gates.
Approval records a decision about a proposal. Authorization controls access. Execution evidence shows what the destination recorded. Keep all three connected and distinct.
A human clicks “Approve,” an agent makes a change and the activity log records success. That sequence can look reassuring while leaving important questions unanswered. What exactly did the human approve? Did the agent still have permission when it acted? Did the destination receive the approved change or a different one?
This article uses three working terms. Approval is a decision about a proposed action. Authorization is the access decision that permits or denies an operation. Execution accountability is the ability to reconstruct what was attempted and what the destination recorded. A reliable workflow connects these facts without treating them as interchangeable.
Make the proposal concrete
Consider a hypothetical operations agent proposing to reduce a staging service from four instances to two. A useful approval view identifies the service, environment, current value, proposed value and reason for the change. It should also explain the relevant consequence in language the reviewer can assess.
A vague request such as “Approve optimization” does not tell the reviewer whether the agent intends to modify staging or production. Nor does it establish which parameter will change. The approval record should preserve the proposal the human actually saw, or a stable reference that allows it to be reconstructed.
This does not mean every action needs a new approval ceremony. The organization can define standing authority for routine, bounded work. The important point is that the action's authority is explicit and inspectable, whether it came from a specific approval or an existing policy.
Enforce permission at the action boundary
OWASP recommends least privilege, denial by default and permission checks on every request. These principles apply to the system that executes an operation; an approval message in a conversation is not itself an access-control mechanism. The OWASP Authorization Cheat Sheet provides the underlying guidance.
In the hypothetical scaling workflow, the destination or trusted execution service should decide whether the acting identity can change that service in that environment. The agent's description of its own authority is not a substitute for that check.
Record the decision and its relevant policy reference. Avoid storing only the current policy if a later reviewer needs to understand an earlier action. The historical question is what authority applied at execution time, not what the permissions screen shows today.
Detect changes between approval and execution
An approved proposal can become stale. The target may change, the desired value may be edited, or the identity's permission may be revoked before execution. Our recommendation is to compare the action being executed with the approved proposal at a trusted boundary.
For the example, a proposal to scale staging to two instances should not silently authorize scaling production to zero. If the material details change, the workflow should follow its defined policy for renewed review or rejection. The comparison must operate on the actual request, not on a later narrative generated by the agent.
Where a workflow uses a digest to identify a proposal, document which fields it covers and how they are serialized. A digest is only useful when the reviewer and executor are referring to the same well-defined content. It does not explain whether the proposed action is a good idea.
Verify the result separately
After execution, retain the destination reference and observed result. If the scaling request was accepted but remains in progress, say that. If the service reached two instances and was later restored to four, preserve both observations.
An approval does not guarantee successful execution. Successful execution does not establish that the approval was valid. A valid permission decision does not establish that the business outcome was beneficial. Keeping these conclusions separate makes it possible to diagnose a failure without rewriting the story around a single green status.
For review, we suggest showing the proposal, approver or policy basis, authorization decision, attempted mutation and destination result together. Each item should retain its own source and time.
Test the uncomfortable cases
Exercise a rejected proposal, an expired approval, a changed target and a revoked permission. Test a valid approval followed by a destination failure. Include a second action in the same conversation so the first approval cannot accidentally be reused for an unrelated request.
Also test the routine path that requires no additional human interaction under the agreed policy. A repair that adds unnecessary gates to already authorized work can make the system less usable without improving the relevant control.
The goal is a workflow in which a reviewer can answer three concrete questions: what was allowed, what was attempted and what was observed. If one answer is missing, show the gap. Accountability comes from preserving those relationships, not from increasing the number of approval buttons.
Sources and further reading
Primary sources checked on 16 November 2026. Examples in this article are illustrative unless stated otherwise.