I gave each of my AI agents a different brain this week. Then I watched one nearly fall asleep mid-task.

I gave each of my AI agents a different brain this week. Then I watched one nearly fall asleep mid-task.

In my homelab I run a small fleet of agents through one gateway. Until now they all shared a single model. This week I routed each agent to the model that fits its job, and instead of buying a bigger GPU, I securely connected to a friend's GPU across the internet.

The orchestrator ("CEO") runs a 120B open model on friend 1's AI host. The two heavy infra agents run on Friend 2's hosted model. Everyone else stays on my local GPU. Call it #meshAI: when you can't afford more silicon, ask a mate. Wire it together over a VPN, and your "cluster" suddenly spans three houses.

The switch itself was beautiful: one field per agent, no redeploys, instantly reversible. So I fired a real task, a quick substrate health briefing, and let the CEO delegate.

Back came one consolidated report: a service-health table from the infra agent (cluster services all active), 11 active VLANs from the network agent, and a clean error scan from observability. One ask in, three brains on three models across three houses, one answer out.

Then the messy bit: that 120B model is 65GB. Cold, the first call took five minutes instead of ninety seconds, long enough that my own watchdog declared the run "stranded" and tried to recover something that was simply still loading.

The lesson: per-agent routing is the easy part. Model readiness is the hard part.

Options I'm weighing: keep the big model warm with a heartbeat, raise the stranded-run timeout, or give the orchestrator a smaller, faster brain and save the heavyweight for the work that needs it.

Right model for the right job, borrowed from the right mate. Including the job of waking up.

#meshAI #HomeLab #AI #SelfHosting