Using an environment

A Savant Environment is a single container image holding one real security defect in one real open-source program, together with everything needed to work on it: the source, a test suite, the tooling to build and run it, and a server that exposes a small, fixed set of operations over the Model Context Protocol.

You supply the agent. We supply the environment, the task, and the grader.

Read the first four pages in order — they take you from an image to a graded episode, and they contain the two mistakes that account for most failed first integrations. Then read the part of Environment types that matches what you were given: the objective and the budget differ substantially between types, and an agent tuned for one will score badly on another.

The rest is reference.