Agent skills written as flow graphs beat free text in the authors' tests

single source· 1 articles · confidence: medium · first seen 2026-09-17 20:00 UTC

What this means for you

If you hand-tune agent prompts, the code is public and the idea is cheap to test yourself. Otherwise nothing to act on: no product, no API, and the gains are measured against a single baseline, SkillOpt, with no evaluation date or harness given.

A paper posted to arXiv on 17 September proposes writing the instructions given to LLM agents as graphs rather than free text: each node is a step with its guidance, and each edge is the condition for moving to the next. The authors then search that structure with a population-based evolutionary method (many candidate skill-graphs kept at once, the better ones mutated and recombined). Across five agent benchmarks it reports average accuracy gains of 4.01% over the baseline SkillOpt on GPT-5.4-nano and 1.76% on GPT-5.4. Code is published; the abstract gives no evaluation dates or harness.

Models in this story

Key facts

  • ·The paper appeared on arXiv on 17 September 2026 under identifier arXiv:2609.21749. source
  • ·Graph-structured skills represent each execution step as a node and context-dependent transitions between steps as directed edges. source
  • ·It reports an average accuracy improvement of 4.01% over the SkillOpt baseline on GPT-5.4-nano and 1.76% on GPT-5.4. source
  • ·Experiments cover five agent benchmarks. source
  • ·Code is released at github.com/ruisun7/GraphSkillEvo. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire