Researchers often face challenges when connecting their Wandb logs to agents for autonomous research loops, due to the slow and clunky nature of Wandb CLI and MCP. This can lead to context rot, hindering the research process.

To address this issue, a new CLI tool and Python SDK, known as Cadenza, have been developed to simplify the connection between Wandb projects and agents. The CLI tool allows for easy import of Wandb projects, structuring runs in a way that enables agents to understand the solution space of the research project.

When projects are imported, Cadenza analyzes and indexes configurations and metrics, storing runs in a way that facilitates efficient sampling by agents. The tool returns only the most high-performing experiments, reducing context rot and allowing for a trade-off between exploration and exploitation.

Cadenza is now open-sourced, along with its Python SDK, making it accessible for use with any agent. The community is invited to provide feedback and critique, with resources available on Github, Docs, and Pypi.

Photo by Artem Podrez on Pexels
Photos provided by Pexels