AI agents that train themselves now have an outside referee checking their work.
Researchers built ANCHOR, a supervisory framework that uses a separate large language model to review self-evolving agents at multiple stages during training rather than only at the end. They added it to two existing open-source self-evolving agent frameworks and tested the results on coding, math reasoning, and safety benchmarks. Safety performance improved substantially while coding and math scores stayed roughly stable. The team also found that grading based on actual execution results worked better than other feedback methods, and that checking in more frequently produced smaller and smaller gains.
Self-evolving agents generate their own practice problems and grade their own answers, which is efficient but leaves nobody watching for slow, cumulative mistakes. The paper describes this failure mode as capability degradation and safety drift. ANCHOR is effectively an admission that letting a model act as its own teacher, editor, and safety officer at once does not hold up, especially as these systems get more autonomy.
Worth remembering: this is benchmark results from one paper across two frameworks, not a shipped product, and substantially safer is measured against the authors' own tests, not an outside standard.