Five Agents on a Mac Mini: What We Learned
Our first PACT5 deployment ran on a Mac Mini. Five agents. One machine. No cloud orchestration layer. No Kubernetes. Just agents coordinating through structured rooms.
We did this deliberately. If the protocol required expensive infrastructure to function, it would fail the practitioners who need it most — small teams running agents locally, testing ideas, building real things without a cloud budget.
Here's what five agents coordinating on commodity hardware taught us.
Structured rooms beat message queues. We started with simple message passing. Agents posted messages, other agents consumed them. It was chaos within a day. Agents talked past each other. Decisions got lost. Conflicts went undetected. When we moved to structured rooms — where agents make formal proposals, discuss, and vote — the coordination quality jumped immediately. The structure isn't overhead. It's the thing that makes multi-agent collaboration actually work.
Agents push back, and that's the point. Our standards agent regularly blocks proposals from the coding agent. Not because it's adversarial. Because it's doing its job — checking whether the proposal meets the criteria it's responsible for. This kind of constructive friction doesn't happen in hierarchical orchestration systems. It only happens when agents are peers in a shared decision process.
Human authority boundaries are non-negotiable. We define exactly where agents can act autonomously and where they need human approval. This isn't a safety feature we added later. It was the first thing we built. Because the moment you give agents real tools, the question isn't "can they do the task?" It's "should they?"
Five agents on a Mac Mini won't impress anyone building at hyperscale. But it proves something important: meaningful multi-agent governance doesn't require massive infrastructure. It requires the right protocol.