How to use the Agent Launcher Orchestrator agent skill

What the skill makes the agent do

Every session starts with one goal: one sentence describing one Claude Managed Agent. The skill reads that goal from ./my-agent/goal.json in your project folder and routes the work to one of five phases: interview, stage-launch, grade-iterate, run-without-you and wrap-up.

The routing is done by a script (goal_router.py), not by guesswork, and a second script (loop_compiler.py) shapes the goal and phase into an execution shape. The three shapes are a single-pass workflow, a bounded grade-to-iterate loop, and a recurring cron deployment loop. Heavy intake such as build sheets, payloads and eval cases is handled in a forked context so it stays out of your main conversation.

Install the agent-launcher-orchestrator skill

Download the skill ZIP from the page you are reading. Then choose one of two locations. For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead of Claude Code, you can upload the ZIP under Settings, then Capabilities, then Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with natural requests

You do not need to type a command to start the skill. The description is written for requests such as the ones below, so describe what you want in plain language and the agent will pick the skill up when the request matches.

• build me an agent
• launch this as a managed agent
• run this on a schedule
• grade my agent against a rubric
• set up a nightly worker

Run the routing and compile steps yourself

The skill runs its scripts from your project folder. First it routes the goal. Exit code 0 means route to the named phase sub-skill. Exit code 3 means the skill will ask you one forcing question and then route again. Exit code 4 means the goal is too vague, so you give one sentence naming the one job and it routes again.

To compile the execution shape, the skill uses loop_compiler.py. The example below asks for up to five grade-and-iterate attempts and a nightly run at 09:00 in the Europe/Berlin timezone, with a self-grading outcome nested in each run.

• Exit 0: route to the named phase
• Exit 3: answer one forcing question, then re-route
• Exit 4: give one sentence describing the job, then re-route

Example: a rubric-graded agent and a nightly worker

Suppose your goal is to triage your inbox every morning. You can start the goal file with goal_state.py, then check its status. The skill will ask forcing questions such as what kicks the agent off, and how you would grade a good run. It recommends a rubric of three to five lines based on the output.

When the agent is ready to launch, the skill emits curl commands for you to run with your own key. It does not send them itself.

• Goal sentence: one job, for example triage my inbox every morning
• Trigger: on demand for the first version, schedule as a later upgrade
• Grading: three to five rubric lines tied to the output
• Integrations: a schema-true mock tool first, a real MCP server later

Safety notes: scripts, network and your API key

The skill runs Python scripts. The scripts use only the Python 3 standard library. Per the skill's own notes, no script in the plugin touches the network, and no tool makes API calls. Every launch is emitted as curl that you run yourself with your own ANTHROPIC_API_KEY from the environment. The launch scripts read the key from the environment and do not print it.

Because you run the curl commands, any API usage happens under your own Anthropic account. Read each command before you run it, and only run it when you are sure of the goal and settings.

• Requires Python 3 and curl to run the emitted commands
• Requires an Anthropic account and ANTHROPIC_API_KEY set in the environment
• No network calls from the scripts; you run the launch curl yourself
• Keep your key out of shared files and screenshots

Tips for a clean first run

Keep one job per goal. If your goal covers two jobs, the skill will refuse to route it and suggests two separate folders, such as ./my-agent-one and ./my-agent-two. Start with on-demand runs and add a schedule later, once the goal is working. Use the grade-iterate loop when you have a clear rubric, and keep the iteration cap low at first.

Only attach a memory store if you want later runs to be smarter than the first one. Otherwise skip it. If you are not comfortable with terminal commands and API keys, do not run the launch steps alone.

• One job per goal file
• Start on demand, then schedule
• Keep max iterations small (the allowed range is 1 to 20)
• Skip memory unless later runs should improve