Cline’s “recursive self-improvement” post is easy to overread as science fiction. The more useful read is narrower and more practical: a coding agent used traces and evals to improve the harness it was running in.
The setup started from Terminal-Bench 2.1. Cline says the stock Cline harness with Kimi K3 through OpenRouter scored 69/89, or 77.5%, at a cost of $79. Then a single prompt launched a long-running campaign with GPT-5.6 Sol as the leader model. The run lasted about 17 hours and consumed roughly one billion tokens, split between agent work and repeated eval runs.
The final confirmation run landed at 79/89, or 88.8%, for $49.8.
What did the agent actually change?
The interesting fixes were not benchmark-specific answer hacks. Cline lists ordinary harness improvements:
- represent max reasoning effort correctly;
- retry provider 429s with exponential backoff;
- make loop detection aware that repeated commands can still be making progress;
- keep an async file-mention lookup alive when prompt tokens look like file references;
- guide agents away from broad
pkill -fpatterns that can kill their own harness.
Cline also says invalid runs were excluded, verifier edits were banned, and the final PR went through human review before merging.
That is the important boundary. The agent did not become its own unchecked judge. It became a worker inside a fixed goal, trace ledger, eval harness, and review process.
New Runtime Read
This is a strong example of agent labor moving from “write code from a ticket” toward “operate an improvement campaign.”
The scarce product layer is not just the model. It is the evaluated loop: baseline, traces, hypotheses, patch, rerun, exclusion rules, cost accounting, and human approval. Once that loop exists, a model can spend real compute turning messy failures into a reviewed harness change.
For coding-agent platforms, this is probably the direction to watch: not autonomous self-modification in the abstract, but long-running, auditable harness maintenance.