spawn

spawn amp agents in tmux windows with automatic thread linkage.

when NOT to use

before spawning, ask:

1. •is this an independent task? spawn for side-quests (different scope). don't spawn multiple agents for the SAME task.
2. •is there a single source of truth? if verifiable against one file/spec/query, do it yourself.
3. •will this produce conflicting findings? spawn for parallel work on DIFFERENT sources, not opinions to reconcile.

spawn parallelizes INDEPENDENT work. don't spawn agents to produce conflicting findings you'll need to reconcile.

why spawn?

short threads outperform long threads. see bundled references:

• •references/200k_tokens_is_plenty.md — why agents degrade with context bloat
• •references/amp_owners_manual.md — official amp prompting guidance

key insight: agents get drunk on too many tokens. spawning keeps each thread focused on ONE task with minimal context, then links via read_thread.

usage

scripts/spawn-amp "<task description>"

creates a detached tmux window, pipes the task to amp, links to parent thread via $AMP_CURRENT_THREAD_ID, includes callback instructions with $TMUX_PANE. echoes the agent's name.

agents get auto-assigned fairy names from assets/ (e.g., ted_glimmermoss).

multiple agents

AGENT1=$(scripts/spawn-amp "task 1") && AGENT2=$(scripts/spawn-amp "task 2")

control

tmux select-window -t "$AGENT1"           # switch to
tmux capture-pane -p -t "$AGENT1"         # check output
tmux kill-window -t "$AGENT1"             # stop
tmux list-windows -F '#W'                 # list all

handoff

when context fills up, spawn a successor with full context:

scripts/spawn-amp "HANDOFF: <context summary>. load the coordinate skill for multi-agent work."

coordination

for orchestrating multiple agents with bidirectional communication, load the coordinate skill.