What
A new safety classifier that reviews sub-agent output when control is handed back to the main agent, checking for potentially dangerous actions performed by sub-agents.
Details
- When a sub-agent completes, its actions are reviewed against security rules before the main agent acts on the output
- If the classifier flags an issue, the main agent sees: "SECURITY WARNING: This sub-agent performed actions that may violate security policy"
- If the classifier is unavailable, a warning note is shown instead of blocking
- Works in conjunction with the auto mode classifier
Evidence
Handoff classifier (search for "Handoff classifier flagged sub-agent output")
Strings lifted out of the shipped bundle, so the claim above can be checked against them.