Some steps take a while. Initialization and generation scale with project complexity; optimization scales with the rollout budget and how long your agent takes to run.
Clone the sample
Creates a disposable local sample workspace.# default location
relai onboarding
# or choose where it lands
relai onboarding --destination ~/relai-onboarding-agent
Enter the folder & set keys
Move in, create the local .env, set OPENAI_API_KEY so the agent can make model calls, then export into your shell.cd ~/relai-onboarding-agent
cp .env.example .env # then edit OPENAI_API_KEY
export $(xargs < .env)
Initialize
Register the sample, write .relai config, and scaffold the simulator. Create a learning environment from the sample log
The feedback tells RELAI what behavior to preserve; a fixed --name makes the generated sample easy to find.relai learning-env create \
--log-file logs/conversation-001.json \
--feedback "The agent should not answer off-topic questions and clearly state it can only help with finance" \
--name no-off-topic
Simulate
Result JSON stays under .relai/runs/.relai simulate \
--learning-envs no-off-topic \
--result-json .relai/runs/onboarding-simulation.json
Optimize
RELAI collects scope, runs improvement attempts, and summarizes proposed changes. Optionally push to a GitHub remote first if you want an optimizer PR.relai optimize --total-rollouts=10
Uploaded runs appear on the Simulations and Optimizations tabs after you open the sample agent from the Agents page.