What's this about?
DarwinX starts from the premise that an agent's capability depends not only on model weights but on its harness — prompts, tools, skills, control flow. Self-improvement loops already edit harnesses, but a single line of edits is path-dependent, and a fix for one task often quietly breaks another. So DarwinX freezes the model and runs natural selection over a *population* of harnesses. A "preserve-and-extend" contract admits only variants that add coverage without regressing existing tasks; an archive keeps alternative lineages available for recombination; and fitness comes from each benchmark's own verifier, so there are no gold solutions and no hand-picked winners. One loop adds about 17 points on average: Terminal-Bench 2.1 reaches 84.7%, WebArena-Infinity real-task pass@1 goes from 43.5% to 93.0%, and a harness evolved on Terminal-Bench transfers unchanged to SWE-bench Verified.
What exactly does "the model is frozen" mean here?
It means no training happens — the weights never change. Everything that improves is the scaffolding around the model. This matters because retraining is enormously expensive and requires access most people don't have, whereas harness changes are cheap text and code. The paper's tagline captures it: "a frozen model need not be a fixed agent." Evaluation compute, normally spent measuring, becomes the mechanism that produces capability.
Why is this worth noticing?
- The transfer result is the important one. A harness evolved on terminal tasks working unchanged on SWE-bench suggests what evolved is general agent competence, not benchmark-specific patches — the usual failure mode of automated optimization.
- The no-regression contract is the mechanism worth stealing. Most self-improvement loops accept any change that improves the target metric; requiring coverage to extend without regression is what stops the search wandering.
- It converges with two other findings this fortnight. OpenART showed the harness explains much of agent safety variation; DeepSeek open-sourced its harness as fully pluggable; DarwinX shows the harness is where cheap capability gains are. The harness is becoming the object of study, not the plumbing.