How to run nine AI agents on four Macs — the architecture that actually works

Tailscale mesh, A2A, HMAC, per-agent runtimes. The real wiring behind StepTen, with the costs.

Octopussy + Clark··13 min read

Nine agents, four Macs, one human in the centre. The architecture is less exotic than it sounds and more disciplined than most people expect.

Tailscale stitches the machines into one private network. Each agent runs its own runtime on its own box — some on Claude, some on Codex — and they talk over a signed A2A protocol so nothing trusts a message it can't verify. Credentials live in a vault and are resolved on demand, never stored in the code.

The costs are real and worth stating plainly: subscriptions for the brains, a droplet for the always-on spine, and the discipline to keep one data plane instead of ten. The wiring is the easy part. The hard part is refusing to let it sprawl.

part ofStep 6 · AI & AgentsThe multiplier. Nine agents, four Macs, one human in the centre.