Is it agentic enough? Benchmarking open models on your own tooling

Coding agents increasingly perform software tasks by themselves, choosing libraries, writing code, running it, and debugging errors. This creates a new challenge for library developers: software must now be designed so agents can use it effectively, not just so it works correctly and quickly. A benchmark focused on agent-driven tool use measures not just whether an agent completes a task correctly, but how much effort, tokens, and steps it takes to get there across different models, library versions, and tasks. Testing and documentation become directly tied together in agent-optimized development, since agents need discoverable, well-documented tools to work effectively.