| name | coordinate |
| description | orchestrate multiple amp agents with bidirectional tmux communication. use for multi-hour autonomous runs with INDEPENDENT parallel tasks. NOT for review courts or generating opinions to reconcile. |
coordinate
orchestrate multiple amp agents. you are the coordinator — agents report to you, you delegate and unblock.
when NOT to use
before coordinating multiple agents, ask:
- is there a single source of truth? if verifiable against one file/spec/query, do it yourself.
- will agents produce conflicting findings? if task is evaluative (judging claims), a single careful pass is cleaner than reconciling disagreements.
- do i have explicit exit criteria? multi-agent work without convergence criteria produces unbounded reconciliation work.
coordinate is for parallelizing INDEPENDENT work. don't spawn review courts when you can read the code yourself.
spawn
use the spawn skill:
AGENT=$(../spawn/scripts/spawn-amp "<task>")
messaging
agent → coordinator (via pane id from spawn instructions):
tmux send-keys -t %5 'AGENT $NAME: <message>' C-m
coordinator → agent:
tmux send-keys -t $AGENT "COORDINATOR: <instruction>" C-m
direct send-keys is preferred. slash commands (like /queue) are unreliable over tmux — timing issues cause messages to be cut off or missed.
monitoring
tmux capture-pane -p -t $AGENT | tail -30
ALWAYS capture before any disruptive action (messaging, killing). never act blind.
control
tmux list-windows -F '#W'
tmux kill-window -t $AGENT
pitfalls
- slash commands unreliable over tmux — use direct messages instead
- permission prompts require arrow keys + enter. ask user to handle manually
- agents can't run sudo with password. have user run, then verify
send-keys -t name targets window name (agents), -t %4 targets pane id (coordinator callback)
