Detecting When an Agent Is Stuck and Should Hand Off
Stuck agents waste budget and erode trust. The handoff signals you can detect cheaply, the handoff destination (human / specialist agent), and the context to pass.
Signals to detect
Iteration count high, progress low. The agent has run 8 steps and is no closer to a hypothesis. Hand off.
Confidence low after multiple investigative rounds. The agent has gathered evidence and still cannot rank hypotheses. Hand off.
Action would be irreversible with limited prior approval context. Even if the agent is confident, the irreversibility is a hand-off trigger.
How to detect cheaply
Counters in the loop: steps taken, hypotheses considered, max-confidence-so-far. The numbers are tracked per run; thresholds fire the hand-off.
Pattern matching on agent state: if the same hypothesis is being investigated for the third time without resolution, hand off.
Self-report: ask the agent every few steps "are you stuck?" The model often knows; let it tell you.
Hand off destination
Default destination: the human on-call for the affected service. They have the missing context the agent lacks.
Alternative destination: a specialist agent. "This looks like a database problem; hand off to the database-specialist agent."
Pick the destination based on the failure signal, not based on default routing. Wrong destination is worse than direct escalation.
Context to pass on hand-off
What the agent tried: the steps, the hypotheses ruled out, the tool calls made.
What the agent observed: key data points, anomalies, things that did not add up.
What the agent recommends: "my best current hypothesis is X with confidence 0.6; I would investigate Y next if I had the tools."
Eval cases for hand-off behaviour
Cases that should result in a hand-off. Pass if the agent hands off; fail if it continues.
Cases that should NOT hand off. Pass if the agent completes; fail if it hands off unnecessarily.
Tune the hand-off threshold to balance: too eager hand-offs erode trust in the agent; too slow hand-offs frustrate operators.