Most autonomous LLM training systems search over training recipes while keeping the evaluation and diagnostic harness fixed. EvoTrainer instead treats the harness as an evolving object: a trainer agent diagnoses rollout-level evidence, revises diagnostics, backtests interventions, and accumulates reusable skills across versions.
In experiments on mathematical reasoning, competitive-programming code generation, and repository-level software engineering, the co-evolved system matched or exceeded human-engineered RL reference pipelines under the same data, codebase, and evaluation protocol. The largest gain was on long-horizon agentic SWE.
The interesting result is that successful strategies diverge across domains, evolving diagnostics prevent invalid high-scoring branches from being promoted, and reusable skills shape later search. This points to autonomous RL that evolves the training harness alongside the policy rather than treating the harness as fixed.