What metric actually matters for coding-agent loops?

[Keesan12]

query: What metric are people actually using for coding agents right now?

The pattern I keep seeing is teams optimize one local number and miss the loop.

Examples:

• cheaper retrieval
• faster tool dispatch
• smaller prompt footprint
• fewer raw tokens in one step

Those all help, but they do not answer the operational question:

did the next attempt deserve to run?

The metric that seems to hold up better is cost per verified outcome:

• what did the run cost
• what verifier admitted success
• what changed between attempts
• why did the loop stop

That is the seam MartinLoop is built around:

• hard budget caps
• verifier gates
• explicit halt reasons
• JSONL run records around the attempt cycle

If you're running Claude Code, Codex, OpenCode, or your own loop, I'm curious what you're actually measuring today:

• raw tokens
• wall-clock time
• task success rate
• human review burden
• something else

If you want to kick the tires, the OSS repo is here:
https://github.com/Keesan12/martin-loop

If that runtime layer is useful, a star helps a lot right now. If you try it, I care more about what breaks.