iterating-to-completion

To prevent premature exits and infinite loops, you can use completion promises and loop detection to monitor agent outputs across planning, implementation, testing, and review phases. This ensures progress until success criteria are met.

The scratchpad protocol maintains a per-task markdown file tracking Completed Steps, Current Focus, Failures & Learnings, and Next Step. This structured logging guides the agent to iterate tasks until all completion signals are observed.

The best way to detect repetitive stagnation is monitoring recent agent outputs to detect repetition, triggering strategy changes or escalation when similarity exceeds predefined thresholds to break the loop.

You enforce explicit completion signals by having agents output specific completion tokens ONLY when criteria are truly met, covering architecture planning, implementation, review, and test passing phases.

Yes, workflow guardrails like iteration caps and explicit orchestrator signals are needed for multi-step agent tasks to prevent uncontrolled looping and ensure successful task completion.