Agent Runaway Task Detection for Coding Agents

Best-fit situations

• An agent repeats the same search query or package installation after each retry.
• A task gets longer while the resulting diff stays empty or unrelated.
• Tests fail repeatedly and the agent keeps patching symptoms without new evidence.
• Tool call volume rises faster than useful commits, PR comments, or reports.

Operating steps

1. Define normal tool call frequency for research, coding, testing, and deployment tasks.
2. Cluster repeated search, install, shell, test, and file-read actions by task.
3. Compare token burn against artifacts such as diffs, passing tests, summaries, and issue updates.
4. Trigger alerts when the run crosses a time, token, or repetition threshold.
5. Recommend safe intervention: reduce model, stop tools, trim context, or hand the task to a human.

Common risks

• A simple time limit misses expensive short bursts of repeated tool calls.
• A simple token limit misses low-cost loops that block delivery for hours.
• False positives can annoy engineers if alerts do not explain the repeated pattern.
• Stopping every long task can prevent legitimate migrations or deep debugging.