Skip to main content
This runs the learning loop end-to-end on your own agent project — initialize, create one learning environment, simulate, then optimize. It assumes the CLI is installed and configured — if not, start with Install & setup. No project of your own yet? Try the sample agent to walk the same loop in your browser.

​
Project requirements

Before running relai init, consider reviewing the compatibility requirements. They cover supported languages, runtimes, dependency declarations, and the adapter capabilities that init can verify locally.
Some steps can take a while. Initialization and generation scale with the complexity of the agent project; optimization scales with the rollout budget and how long the agent takes to run.
1

Optional — limit project inputs

Add a root-level .relaiignore file for files and directories the CLI can ignore because they are not needed to import, configure, or simulate your agent. RELAI uses these patterns to keep matching project files out of agentic init model context and other generation snapshots. Replace placeholders with project-specific paths. See Controlling project inputs with .relaiignore for scope and examples.
2

Initialize the repository

Run from the root of your git-tracked agent project. RELAI seeds and validates the simulator harness under .relai/simulator/, then registers the agent. If init pauses for an answer, approval, external action, or interruption, continue it with relai init --resume.
3

Create a learning environment

From a prompt, or from a run log plus feedback. The input log can live anywhere — point --log-file at your own path. The generated environment always lands in .relai/learning-envs/.
4

Optional — register a benchmark

Reuse the same evaluation suite across simulation and optimization.
5

Simulate behavior

6

Optimize

Improves the agent against selected environments and benchmarks. When done, it may write changes in a separate worktree and open a GitHub PR if GitHub CLI (gh) is available.

The learning loop

The loop and the objects it uses, step by step.

Command reference

All flags, defaults, and failure modes for every command.