EvoSkill v1.3.0 Runs Self-Improving Coding Agents Without Claude API

0xsachi · x · 2026-08-01

Sentient's open-source framework EvoSkill shipped v1.3.0, adding Fireworks AI as a first-class inference provider. This allows developers to run the entire agent self-improvement loop—including the evolution harness and LLM scorer—on fast, cost-effective open-source models without needing the Claude API.

EvoSkill continuously improves itself by analyzing failure cases to automatically generate structured skills. Benchmarks indicate a 7.3% improvement on OfficeQA and 12.1% on SealQA, alongside a +5.3% zero-shot skill transfer to BrowseComp.

Original post →

More from coding & agent

coding & agent channel →