Automating the design of language-model agents has been a step forward, but the process still rests on a human-built scaffold. According to a new arXiv preprint, agent evolvers can generate prompts, skills, and workflows on their own, yet the optimization procedure itself remains a fixed loop that researchers must specify in advance. That loop determines how candidates are evaluated and, presumably, how the search progresses.
FreeEvolve, the method introduced in the paper, is described as learning to evolve beyond those fixed loops. Rather than taking the search loop as a given, it appears to treat the loop itself as something to be learned or adapted. The abstract is brief, but the core claim is clear: the next frontier in agent automation may be automating the optimizer rather than only the artifacts it produces.
Because the source is a single preprint, the details of how FreeEvolve works and how well it performs are not yet available. The paper does not claim to have solved the problem, only to move beyond the fixed-loop assumption. As with most arXiv announcements, the ideas are preliminary and will need further validation before they can be judged against existing methods.