Agent Collective
Agent-to-Agent Messaging

Challenge
Running several agent sessions in parallel, one per repo or service, leaves each of them blind to the others. Passing context between them meant copying and pasting by hand.
Messaging also had to stay out of the way: no announcements when sessions start or stop, nothing added to prompts, no interrupting a busy session, and addresses that survive restarts and resumes.
Results
• Two agent tools: peers lists who is online, and peer_send prompts a session by callsign, with the reply arriving as a peer message.
• Scoped by omp profile: sessions in one profile never see sessions in another.
• Silent and need-to-know: nothing is broadcast, and a session’s contacts are only the peers it has messaged or heard from.
• Safe delivery: an idle peer wakes into a turn, while a busy omp peer receives the message as an aside without interrupting its running tools.
• Stable callsigns: numbered per directory, reserved for 7 days after a session closes, and claimed with exclusive file locks so no two sessions share a name.
• /collective terminal UI: a bottom panel, a full-screen list with a detail column and per-peer threads, plus a CLI and a Tern plugin.













