Agent management

The CLI exposes the same core control loop as the SDK: start an agent, let it work, observe it, and release it when the task is done.

Spawn an agent

agent-relay spawn reviewer claude "Review the latest auth changes"

Useful options:

• --cwd <path> runs the agent in a specific working directory.
• --team <team> tags the agent with a team name.
• --model <model> overrides the default model for that CLI.
• --continue or --continue-from <name> resumes a previously released agent identity.
• --shadow-mode <mode> and --shadow-of <name> configure shadow execution for advanced setups.

You can also pipe the task through stdin:

printf '%s\n' "Review the auth diff and list regressions." | agent-relay spawn reviewer claude

Inspect active agents

agent-relay who
agent-relay agents:logs reviewer --follow

• who shows agents active in the last 30 seconds.
• agents:logs tails stdout and stderr from a spawned worker.

Release or retarget an agent

agent-relay release reviewer
agent-relay set-model reviewer sonnet

• release shuts down a spawned agent cleanly through the broker API.
• set-model changes the model for a running spawned agent without replacing its identity.

Practical pattern

agent-relay up --background
agent-relay spawn planner claude "Break the task into three steps"
agent-relay spawn coder codex "Implement the approved plan"
agent-relay send planner "Coordinate with coder and keep updates concise."

See also