Skip to main content
Debug Mode pauses operations that policy would have allowed, so you can watch what your agent actually does. Most teams are surprised the first time.
Or from the operator view, or the control plane:

​
What you see

For each paused operation: the statement as it will run — scoped, column-restricted, masked — its parameters, its class, whether it is provably bounded, how many records it estimates, the hazards it carries, and the agent’s own stated reason for doing it. For a change, the same preview a reviewer gets.

​
What you can do

Ending a session refuses what is pending rather than letting it through: a session that ends because you closed your laptop should not commit a queue of changes.

​
Filters

Pause everything, or only what you care about: Sessions can also be limited to particular agents, and carry a TTL so one left open expires on its own.

​
Exporting a session

The export is every operation the session saw, with your verdict attached to each: the statement, the resources, whether it was bounded, what you decided, and why. That is a labelled record of what an agent tried to do with a human’s judgement beside it — directly usable as an evaluation set for improving the agent.
Debug Mode always blocks, even when the connection or the request asks for async approval. It is an interactive tool: a person is stepping through the agent live, and there is nothing to poll for.