Add or edit Guards
Updated October 19, 2026 · 2 min read
Add Guard rules on a worker so the final result can't skip the policy you care about. A Guard sees that result plus the Worker Actions taken in that run (id, status, short summary). It doesn't see conversation history, knowledge, Skills, or other tool traces.
Open Guards
- Open a worker, go to its Guards tab, and expand Configure guard rules.
Add or edit a rule
- Click Add guard. Write the criterion so it can be judged from the final result plus the Worker Actions taken in that run, add optional exceptions, and save.
- To change an existing rule, open it, edit the criterion, and save again.
What a Guard can see
A Guard sees the worker's final result plus a short list of the Worker Actions taken in that run (action id, status, short summary). It doesn't see conversation history, knowledge, Skills, or other tool traces. Write the criterion and exceptions so they can be judged from that result and those Actions.
Criteria that work, and ones that don't
Works | Doesn't work |
|---|---|
The reply promises a refund. | The reply contradicts what the customer said earlier. |
The reply says the case was escalated, but no escalation Action succeeded this run. | The reply ignores our help center article on pricing. |
A Slack message Action was sent, but the reply doesn't tell the customer a teammate will follow up. | The worker used the wrong Skill. |
FAQ
What happens when a Guard is broken?
The draft that broke the rule is blocked from going out. Broken guards show up on the Result row so you can see why.
Do Guards apply to every Skill?
Yes. Guards apply worker-wide, regardless of which Skill produced the result. They see the final result plus the Worker Actions taken in that run, not which Skill ran, the knowledge it used, or earlier messages.
Can I see which Guard blocked a reply?
Yes. Open the Result row — broken guards are listed there alongside the draft that triggered them.
Can I delete every Guard on Chat or Internal Q&A?
Yes. Remove them on the Guards tab and save. An empty list stays empty — built-in presets do not come back unless you add them again from Available guards.
Can a Guard check whether an Action actually ran?
Yes. The Guard sees each Worker Action from this run with its status and a short summary, so a criterion like "the reply says it was escalated but no escalation Action succeeded" can be judged.
Related articles
- Add an action to a worker
Add actions to a worker so every Skill inherits them, or to one Skill only, like Slack messages, Crisp updates, or follow-ups.
- Add a trigger to a worker
Add a live or scheduled trigger so a worker only runs on the events you choose.
- Edit Agent Skills for a Play
Open Knowledge → Agent skills to edit the playbook a Play reads at runtime.
- Schedule a follow-up from a worker
Let a worker schedule a follow-up for itself, then come back later to check whether a fix worked or a teammate answered.