← News

Agents & MCPSources: arXiv · 19 Nov 2026

GRAFT uses execution evidence to guide evolutionary LLM agent design

The GRAFT paper proposes guiding evolutionary agent design with parent-specific execution traces rather than aggregate fitness scores alone, reporting higher test accuracy across six reasoning and programming settings.

In plain words

Automated agent design uses evolution-like search to build programs that help large language models solve tasks. The GRAFT method keeps the usual fitness-based selection but also feeds the search detailed records of how parent agents succeeded or failed on specific tasks, claiming this produces better agents without extra task runs.

What was announced

Automated agent design searches for executable architectures that improve LLM-based agents, and evolutionary approaches typically recombine promising mechanisms using language-model crossover. The authors of GRAFT argue that this search is usually guided by aggregate fitness alone, which hides the instance-level successes and failures behind each score. GRAFT instead uses aggregate fitness for selection and survival, while parent-specific execution evidence guides a semantic crossover step with integrated mutation.

The feedback harness reuses bounded samples of successful and failed runs from fitness evaluation, withholds reference answers, and requires no additional task executions. In evaluations across six reasoning and programming settings using open-weight optimizer and executor models, the authors report a mean test accuracy of 82.0%, exceeding the strongest baseline by 9.8 percentage points. Controlled ablations, with initialization, search rules, repair policies, and evaluation budgets held fixed, reportedly show that adding execution evidence improves test accuracy. The results support parent-specific execution evidence for evolutionary agent design, though the figures come from the paper's own abstract and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so method details and evaluation conditions cannot be verified.
  • The accuracy figures and ablation findings are the authors' own reported results and do not constitute independent confirmation.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.