How to Reduce Wasted Agent Runs
Remove avoidable ambiguity before delegating substantial work.
Give the agent a usable starting point
Include the correct repository or URL, the intended deliverable, relevant prior progress, and what must remain unchanged. Missing context can lead to rebuilding the wrong thing or requesting information you already have.
Define a stopping point
“Improve until perfect” has no clear end. Acceptance criteria let the agent stop when the requested outcome is met. Resource strategy should guide depth without pretending to predict tokens or usage.
Keep validation proportional
For a focused bug fix, validate the affected behavior and concrete regression risks. For a consequential analysis, check source quality and disconfirming evidence. More activity is useful only when it resolves relevant uncertainty.
Preserve progress
Ask for checkpoints when a project is substantial. If a tool fails, use a bounded alternative and continue independent work. Record what is complete and the next step so another run can resume without restarting.