Users will want to stop an agent mid-run, correct a misunderstanding, or undo something it did. All three are harder with agents than with normal features, and all three are load-bearing for trust.
Capability | What It Needs | Commonly Missing |
|---|---|---|
Stop | Halting cleanly and keeping partial findings | Stop that discards everything it learned |
Correct | Feeding a correction in and continuing | Having to restart the whole task |
Undo | Reversing reversible actions individually | No record of which actions were taken |
Take over | Handing the task to the human with full context | Handover that loses the trace |
The most common implementation of 'stop' throws away everything the agent had found. From the user's side that means interrupting costs them forty seconds of progress, so they don't interrupt, they watch it go wrong to completion. A stop that preserves findings makes interruption cheap, which makes users willing to delegate in the first place.