Clear objective
Tell the agent exactly what job it owns and what success looks like.
MartinLoop keeps the job reliable from the moment you hand it off until you get an outcome back. The agent still codes. MartinLoop handles the boundaries, checks, recovery, and record around the work. Start with the guide to Ralph loops, Claude Code loops & Codex loops, or watch the live MartinLoop vs Ralph Loop demo.
Tell the agent exactly what job it owns and what success looks like.
Keep one task from quietly turning into a rewrite.
Give the run hard limits on spend, time, attempts, or allowed paths.
MartinLoop can end runs that hit configured limits instead of letting them grind forever.
Your configured tests, lint, typecheck, or other verifier decides whether the evidence passed.
Every run closes as VERIFIED, STOPPED, or NEEDS REVIEW, with a clear stop reason.
Keep enough evidence and state to repair, retry, roll back, or hand the job to a person.
See the task, attempts, checks, outcome, and cost together instead of reconstructing the run yourself.
See what finished, what failed, what needs attention, and what the work cost. Atlas records the run, SANSA trace intelligence explains evidence-backed failure patterns, and governed learning surfaces evaluated improvement evidence without silently changing production behavior.
Read the FAQ · view the roadmap · explore the agent loops guide