Roll out agent controls in three phases. Audit mode records what would be blocked without stopping anything. Warn mode tells the user and the agent owner. Block mode refuses the action. Move up only when the false alarm rate is low and owners know the rules. Set an exit test for each phase so none drags on.
The three modes
| Mode | What happens | Purpose | Typical length |
|---|---|---|---|
| Audit | Action allowed, event recorded | Learn real behaviour, tune rules | One to two weeks |
| Warn | Action allowed, user sees why it is risky | Teach, collect objections | Two to four weeks |
| Block | Action refused with a clear reason | Enforce | Ongoing |
The lengths are starting points. Your usage volume decides the real numbers.
Phase one: audit
Turn the rules on in record-only mode across a small, willing group. Review every event that would have been blocked. For each one, decide whether it is a true problem, a rule that is too broad, or a legitimate need the rules lack. Fix the rules, not the people.
Exit test: for one full week, almost all would-be blocks are things you agree should be blocked, and there is no unexplained noise.
Phase two: warn
Show the user a short message: what was risky, which rule, and how to request an exception. Count how many warnings people act on. If they ignore most, either the rule is wrong or the message is unclear.
Exit test: owners understand the rules, the exception route works, and a sample of warnings has been reviewed by a person.
Phase three: block, selectively
Do not block everything at once. Start with the actions where harm is highest and the rule is clearest, such as deleting data, sending outside messages or changing permissions. OWASP lists requiring human approval for high-impact actions among its measures against excessive agency, and a block with an approval path fits that.
Keep lower-risk rules in warn mode until you have the evidence.
Run an exceptions route
People will hit a block and need an answer. Provide a form, a named reviewer, and a target response time. Log each exception with a reason and an expiry date, and review them monthly.
Watch the numbers
Track events per mode, percentage that were false alarms, exceptions requested and granted, and time to resolve an exception. If the false alarm rate stays high, stay in the earlier phase. The related post on shadow mode for automated remediation uses the same idea for fixes.
Try it yourself
Click through it at your own pace, or open it full screen.
See how ongoing monitoring surfaces events over time, which is what you review during the audit phase.
Safeguard policies, gates and guardrails can be tuned stage by stage, and the free scan is a gentle first step.
Quick check
Can we skip audit mode?
You can, but you will learn your rules are wrong from angry users instead of from data.
Should every team move at the same speed?
No. Teams with simpler agents can move sooner. Keep the exit tests the same for everyone.
Sources
Last reviewed: 19 October 2026.
