Crew hands work off. You don't shuttle messages.
The handoff bus is the message floor where agents pass work to each other — envelopes, results, leave-behinds, and aggregated lane outcomes — without routing through you.
Your agents pass work to each other directly — no babysitting copy-paste between tabs.
Typed envelopes in local SQLite (WAL). Recipients claim atomically with BEGIN IMMEDIATE so two workers can't grab the same job; TTL expires inline during claim — no background daemon to die and leave zombie work.
dig in
BEGIN IMMEDIATE; -- one recipient wins the claim; the other backs off
The bus is local to one box — agents share a single SQLite workspace, not a network. You can run that box anywhere: your laptop, or a VPS you reach over SSH or Tailscale (we smoke-tested VPS-from-day-one). What it's NOT yet is a mesh of agents coordinating across separate machines — that's roadmap, not shipped. We'd rather nail one box bulletproof than ship a distributed bus that silently drops messages.
You can. You'll have the happy path running in a weekend. Then two agents claim the same ticket because you skipped the atomic lock, your retry daemon dies overnight and leaves work in limbo, and month two is you rebuilding a message bus instead of your product. We already paid that tax — this is the part where DIY buys you nothing but lost time.
