Skip to main content
RELAI works best with Python agent projects that can run locally and are committed to git.

Compatible projects

Before running relai init, make sure that:
  • the project is a git repository with at least one commit.
  • the RELAI simulator virtual environment can run Python 3.11 or newer.
  • the project has a clear agent entrypoint, function, command, graph, or framework object.
  • dependencies are declared in files such as pyproject.toml, requirements.txt, uv.lock, poetry.lock, setup.py, setup.cfg, Pipfile, pdm.lock, or Conda environment files.
  • runtime secrets come from environment variables or local env files, not committed source.
  • the behavior you want to test can run without a production-only hosted boundary.

Frameworks

RELAI aims to be framework-agnostic: the core loop works with any Python agent that exposes callable components — functions, tools, a graph, or a clear entrypoint — so you don’t have to adopt a specific library to use it. Support is deepest today for OpenAI Agents SDK, LangChain, and LangGraph, which have explicit, tested integrations. Other frameworks work on a best-effort basis as long as they expose callable Python components — coverage there is still a work in progress, and we’re actively expanding it.

Limits on file sizes and arguments

ItemLimit
Learning environment prompt8 KB
Learning environment feedback8 KB
Learning environment log file48 KB
Evaluator prompt8 KB
Markdown memory file16 KB
Simulation artifact upload file100 KB
Selected learning environments per simulate or optimize invocation1000
Optimizer --total-rollouts1000
Optimizer --batch-size3