Posts by the Skillberry team, collected from wherever they were published. Newest first.
With its weights unchanged, Gemma-4-31B goes from 43.5% to 76.8% on held-out SpreadsheetBench tasks once cap-evolve rewrites its instructions. That is 8.8 points above the best published result. The post covers the method and how the comparison was kept fair.
Tests a model writes from a single prompt are a vibe check with checkmarks. This post builds a lab for agents (Simulation Harness for the world, Rubrica for trusted tasks, cap-evolve for the fixes) and uses it to find real bugs in Red Hat's PARSEC.
A higher score doesn't tell you why. The new CapEvolve UI traces each failing task to its root cause, the edit that fixed it and the candidates the gate refused. In one run, a single prompt rule lifted held-out accuracy from 38.6% to 89.8%.
Audits found over a third of published agent skills carry a security flaw. Skillberry-Store is one Apache-2.0 service that versions, governs and serves every skill to your agents, with plugins that scan and approve skills before they reach a runtime.
Some agent failures are structural, and prompt tuning won't fix them. Meet cap-evolve: an optimization loop over an agent's prompts, tool code, MCP surfaces and skill packages, driven by an agent, on ฯยฒ-bench, SWE-bench and SkillsBench.
