Governance You Can Show Someone Else
Controls only count if you can show them working. Three records answer the questions a reviewer asks: who changed what, who approved it, and what the agent actually did.
In this post
- 01The question after the incident
- 02Three records, three questions
- 03What changed, and who changed it
- 04Who approved it
- 05What the agent did
- 06You choose how much to keep
- 07A Monday morning, answered in minutes
- 08What reviewers ask for
- 09Evidence is not compliance
- 10A review you can run in an hour
- 11Your first week
The question after the incident
Every incident review reaches the same three questions, and they are rarely asked kindly.
Picture this
A model bill doubles over a weekend. On Monday a manager asks: what changed, who allowed it, and what did the agent do? If the answers live in three people's memories and a chat thread, the review turns into a debate. If they live in records, it turns into a fix.
Having controls is the first half of governance. Being able to show them working is the second.
Three records, three questions
Cloptima keeps one record for each question.
| Question | Record | Where |
|---|---|---|
| What changed, and who changed it? | The Control Plane Audit Log | AI → Audit |
| Who approved the risky ones? | The Approval queue and its decisions | AI → Audit |
| What did the agent actually do? | Session traces | AI → Sessions |
Each is written as things happen, not reconstructed afterwards.
What changed, and who changed it
Changes to your gateway are recorded with the person, the time, and the target.
- Policies and bindings: created, updated, deleted
- Virtual keys: created, updated, rotated, revoked
- Provider credentials: created, validated, rotated, revoked
- Tool servers, guardrail profiles, edge instances, and cache invalidations
You can filter by action, search by name, and export the log for a review.
Who approved it
Some changes can raise cost or risk, so they ask for approval before they take effect.
| Change | Asks for approval |
|---|---|
| Raising a budget | Yes |
| Allowing a costlier model | Yes |
| Activating a tool server | Yes |
| Switching the semantic cache to Enforce | Yes |
| Lowering a budget or removing a model | Never |
A reviewer approves or rejects with a note. Someone below admin cannot clear their own request. An admin who needs to move fast can apply a change immediately, and the self-approval is recorded.
What the agent did
Session traces turn an agent run into a readable sequence.
- Every request in the session, in order, with model, cost, and status
- The tools it called, with arguments and results, if you choose to keep them
- A verification state on each record, so you know it is intact
A log you cannot verify is a story. A log with a fingerprint is evidence.
You choose how much to keep
Privacy and evidence pull in opposite directions.
The policy lets you choose where to stand.
| Mode | Keeps |
|---|---|
| Zero retention | Counts and cost only. The default |
| Attribution only | Labels, metrics, and tool names |
| Tool calls | Tool arguments and results |
| Full audit | The text of every step |
A Monday morning, answered in minutes
Return to the doubled bill. With the three records, the review goes like this.
- 1
What changed
The Control Plane Audit Log shows a policy update on Friday at 17:12 that raised the daily budget and added a model.
- 2
Who approved it
The Approval queue shows the request and an admin's approval with a note: temporary for a launch.
- 3
What the agent did
The session view shows one run that repeated the same tool call sixty times over the weekend.
Three different causes were in play: a loosened limit, an approved change, and a looping agent. Without the records they blur into one argument. With them, each gets its own fix.
What reviewers ask for
Auditors and security reviewers ask for the same kinds of evidence again and again.
| They ask | You show |
|---|---|
| Who can change AI settings? | Owner and admin roles, and the audit log of their changes |
| How are risky changes controlled? | The approval types and the decisions in the queue |
| Can you reconstruct an agent's actions? | A session trace with a verified record |
| Is sensitive data kept? | Your retention mode, and the guardrails that run before storage |
| How do you respond to an incident? | The path from Explorer to Audit to Sessions |
Evidence is not compliance
Be honest about what records prove.
These records help you document and run your own controls. They do not make you compliant with any standard on their own. They give your auditors something concrete to examine, which is usually where an audit gets easier.
A review you can run in an hour
A short, regular review keeps the records useful.
- 1
Approval queue
Clear anything pending, with a reason.
- 2
Policy Violations
Read the top groups and fix any that block real work.
- 3
Control Plane Audit Log
Scan policy, key, and credential changes for surprises.
- 4
Sessions
Open the most expensive agent session and read it.
- 5
Export
Save the log for the month.
Your first week
Start with the records you can turn on today.
- Open the Audit tab and read each card once
- Name your reviewers for risky changes
- Choose Tool calls retention for one agent policy
- Put the Audit tab on your on-call checklist
- Run the review above at the end of the week