OpenHands Agent Loop: run() and step() Diagram

Loop ownership

LocalConversation owns the while-loop, the FIFO state lock, statuses, max_iteration_per_run and max_budget_per_run
Agent.step is one iteration; it never loops and keeps no history of its own
Every effect leaves step() via on_event, which persists before notifying subscribers
Before each step: STUCK if StuckDetector sees repeated action/observation, error or monologue patterns

Early returns

A Condensation or CondensationRequest ends the step; the next iteration reads the condensed view
Malformed tool calls and content-filter blocks become user-role nudges, not crashes
Confirmation mode stops after ActionEvents are recorded; the next run() executes them

Next iteration

Without finish, the observations are in the log and run() calls step() again
A plain text answer sets FINISHED; a new send_message resets FINISHED to IDLE