An interruption is a prompt you send while the agent is still working — before it finished the previous turn. A few are normal. A high rate is usually a signal that your opening prompt was under-specified, so you keep course-correcting mid-flight.

Why it matters

Every interruption throws away in-flight work and forces the agent to re-read context to recover. That burns tokens, breaks the plan it was executing, and tends to produce shallower results because it never gets to finish a thought. The cost compounds: each interrupt usually triggers another tool round-trip to re-orient.

How to improve

Front-load the first prompt. Spend an extra 20 seconds stating the goal, the constraints, and the “done” condition. The clearer the target, the less you’ll need to redirect.
Use plan mode for anything non-trivial. You approve the approach before any work starts, so you redirect plans (cheap) instead of in-progress work (expensive). See use plan mode.
Let it finish, then steer. If the agent is close, let the turn complete and correct in the next prompt rather than cutting it off.
Queue, don’t interrupt. If your tool supports queuing a follow-up, that’s gentler than a hard stop mid-tool.