Shows an agent optimizing its own harness for a model, reaching a reported 88.8% on Terminal-Bench 2.1 for about $50. Engineers can see what automated harness tuning costs and what it can replace.
Key takeaways · AI-distilled
Cline ran a single-prompt campaign of about 17 hours with GPT-5.6-Sol as the leader model, using roughly a billion tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → and $680 in total, versus a January hill climb that took four engineers a couple of weeks of trace reading.
The baseline Kimi K3 run on stock Cline scored 77.5%; Cline says five of those failures were OpenRouter 429 rate limits that killed healthy sessions, fixed by more retries with exponential backoff.
The AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → made Cline's loop detector output-aware, since agents legitimately polling long-running background jobs repeat the same command while the output keeps changing, and the old detector was killing them.
Two tasks failed because the agent ran pkill -f with a pattern matching its own agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → command line; tool guidance to track PIDs instead helped but, per Cline, did not fully eliminate broad-match kills.
Cline says the prompt explicitly banned reward hacking such as verifier edits, task-name detection and timeout inflation, and humans still reviewed the resulting PR before merging.
Terms in this piece · Glossary
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
