Skip to content

Agent Runtime

The lab runs an actual fleet of agents in production — personal assistants, company operations agents, and tenant agents — and treats them as production software: versioned, reviewed, canaried, monitored, and recoverable.

Agent fleet architecture — channels, gateways, router, platform services
Channels in, gateways on k3s, a millisecond router, and durable sagas feeding observability.

What a "gateway" is

Each agent runs in its own gateway pod: the agent runtime, its workspace and state (on replicated storage), its own LiteLLM virtual key, and its own broker token. Gateways are multi-tenant by construction: no shared state, no shared keys, no cross-tenant access paths.

Fleet operations

The fleet is operated like production software: immutable versioned images, canary-first rollouts, versioned read-only baselines, durable sagas with verified postconditions, automated health audits, and authenticated support routing. See Agent fleet operations for the full picture.

Why fleet-grade discipline for a "lab"

The point of the lab is not "an agent that works on my laptop". It is: can agents be operated like production software? The fleet answers yes — and the patterns (immutable images, canary waves, baseline versioning, durable sagas, tenant isolation, automated health) are exactly the platform engineering that AI Platform Engineer roles are about.