cap-evolve
Optimize any AI agent's skills, tools/MCP and prompts against your own evals.
🫐 Open source · Apache-2.0
Capabilities your agents can trust.
Skillberry provides a framework for enterprise-ready agentic capabilities — skills, tools, and prompts. It emphasizes the usage of approved, determinism-rich capabilities at runtime to minimize reliance on LLMs and enhance trustworthiness.
- Skillberry provides scalable skills, tools and MCP runtime management and automation.
- Skillberry performs iterative optimization to accelerate generation, testing, and refinement of agentic capabilities.
- Skillberry generates simulations and benchmarks from real test cases for enabling in-context agentic optimization.
Manage
Skillberry provides scalable skills, tools and MCP runtime management and automation.
Optimize
Skillberry performs iterative optimization to accelerate generation, testing, and refinement of agentic capabilities.
Simulate
Skillberry reliably simulates and benchmarks real-world agentic interactions, built from real test cases.
🤖 Key projects
All projects →skillberry-store
The open-source control plane for agent skills, tools and snippets.
simulation-harness
Simulate MCP tools and skills — no backends, credentials or side effects.
rubrica
Skill-based test suite generator for agentic systems.
llm-switchboard
One client for many LLM providers — tool calls, structured outputs, retries.
skillberry-agent
A proxy-agent that connects your agents to Skillberry skills.
📝 From the blog corner
All posts →
cap-evolve beats the state of the art on SpreadsheetBench: Gemma-4-31B from 43.5% to 76.8%, weights unchanged
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.
Agents in the Lab
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.
Watch Your Agent Evolve: Inside the New CapEvolve Observability UI
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%.
📺 On YouTube
@skillberry-ai →🤖📺 The bots are still in the editing room — no public videos yet.
Subscribe to @skillberry-ai to catch the first upload.