// Use when the user asks to check on agents, manage their fleet, babysit panes, see what needs attention, get a status report, or when they say "check on my agents", "what needs action", "fleet status", "babysit", "brief me". Requires the control MCP server to be registered and the daemon running.
Use when evaluating, designing, or pressure-testing the business model of an AI agent product. Triggers on "agent business model", "agent economics", "agent canvas", "evaluate agent product", "agent pricing", "agent unit economics", "agentic business", "AI agent company", "agent cost structure".
Use when someone wants to launch a content brand from scratch on any platform. Triggers on "launch a channel", "start a YouTube channel", "build a podcast", "start a newsletter", "content brand", "creator launchpad", "build my audience", "start creating content", "grow on TikTok", "start a blog".
Use when you have an existing pipeline DOT file that needs review, fixing, or validation. Use when validation produces warnings or errors, when a DOT file was hand-written or generated by another tool, or when you want to verify a pipeline covers 100% of a spec before running it.
Use when you want to build a project but don't have a spec yet and need to brainstorm the idea into a design doc structured for pipeline DOT generation. Use when starting from a vague idea, project concept, or feature request that needs to become a headless autonomous build pipeline.
Use when you have a spec, design doc, structured plan, or prompt_plan and need to generate a pipeline Graphviz DOT file for headless autonomous building. The agent reads the spec, extracts build pipeline phases, determines tech stack, asks about human gates and model preferences, generates a validated DOT digraph with rich self-contained prompts, and saves it to the project root. Handles any language or framework.
Use when a pipeline DOT file has passed structural validation but may still have flow inefficiencies, spec fidelity gaps, or information loss in prompts. Use after dotfile-audit or dotfile-from-spec when you want to deeply verify a pipeline captures everything from the original spec. Use when prompts might be missing critical domain rules, data models, or invariants.
| name | fleet-manager |
| description | Use when the user asks to check on agents, manage their fleet, babysit panes, see what needs attention, get a status report, or when they say "check on my agents", "what needs action", "fleet status", "babysit", "brief me". Requires the control MCP server to be registered and the daemon running. |
Manage a fleet of AI coding agents running in tmux panes using the control MCP tools.
digraph fleet {
"list_panes" -> "Filter to claude-code only";
"Filter to claude-code only" -> "get_pane_conversation for each";
"get_pane_conversation for each" -> "Classify each pane";
"Classify each pane" -> "Present digest to human";
"Present digest to human" -> "Human picks action";
"Human picks action" -> "Execute (send_to_pane, stop_pane, or skip)";
"Execute (send_to_pane, stop_pane, or skip)" -> "list_panes" [label="next cycle"];
}
Call list_panes with no filters. Then for each claude-code pane, call get_pane_conversation (count: 10) to get recent context.
Read each conversation and classify:
| Status | Signal | What it means |
|---|---|---|
| Working | Recent tool calls, thinking, actively producing output | Leave it alone |
| Needs input | Asked a question, waiting for user response, permission prompt | Human needs to respond |
| Stuck | Repeated errors, retrying same thing, no progress in last few messages | May need redirection |
| Idle | At the prompt, no recent activity, conversation ended | Available for new work |
| Done | Completed its task, committed, reported back | Can be stopped or given new work |
Use your judgment. Read the conversation content — don't just look at the status line.
Format as a brief table:
Fleet Status: 5 claude-code panes
%4 nl Working editing graph layout code
%19 tracker Needs input asking about test strategy
%20 barnstormer Idle at prompt, finished last task
%50 hex Stuck retrying failed SSH connection
%79 dot-files Done committed linting fixes
1 needs your input, 1 looks stuck. Want me to handle any of these?
Keep descriptions to ~5 words. The human wants a glance, not a novel.
When the human says to act on a pane:
send_to_pane with the answersend_to_pane with new instructions ("try a different approach — skip the SSH and test locally instead")send_to_pane with a new taskstop_paneget_pane_conversation with higher count (30-50)When sending to a pane, be direct. The text you send is typed into the pane and Enter is pressed. Write it like you're talking to the agent.
If the human says "handle it" or "you decide":
send_to_pane.Never stop a working agent autonomously. Never send instructions you're unsure about.
| User says | What to do |
|---|---|
| "check on my agents" | Full loop: list, classify, digest |
| "what needs attention" | Same, but only show needs-input and stuck |
| "brief me" | Full digest with one-line summaries |
| "handle the stuck ones" | Read stuck conversations, send redirections |
| "send X to the tracker agent" | send_to_pane using project name fuzzy match |
| "stop the idle ones" | stop_pane for each idle pane |
| "what's %4 doing" | get_pane_conversation for that specific pane |
The control daemon must be running and the MCP server registered. If tools aren't available, tell the user:
The control MCP tools aren't available. Make sure:
1. The daemon is running: control daemon start --background
2. MCP is registered in ~/.claude/settings.json