On the fleet
Pool-based training host selection, deterministic routing, and what a fleet of specialists actually looks like.
A fleet of specialists implies a fleet of training hosts. We treat the hosts as a pool, not as static cluster-to-host assignments. This is the single design decision that makes the autonomous iterate loop survive a hardware outage at three in the morning.
Pool-based host selection
At training time, the iterate engine probes every host for reachability. It scores each by affinity and availability. It selects the best candidate. If a host goes offline, its clusters fail over transparently to the pool. No configuration change. No manual intervention.
This is not sophisticated. It is the minimum bar for a loop that is meant to run continuously. If a Spark drops off the network and the training run dies with it, the system is not autonomous. It is supervised, and the supervisor is exhausted.
Routing is not the moat
Query routing is handled by a deterministic rules engine, not a trained model. This is intentional. Routing is not the moat, the specialists are. A deterministic router is predictable, auditable, and debuggable. It does not introduce an additional trained component whose failure modes cascade across the entire fleet.
Teams that try to make their router clever usually end up with a system that is brittle in proportion to its cleverness. The router grows a long tail of edge cases. The specialists it dispatches to become harder to evaluate because the router itself is now part of the question under test. The moat dilutes across two trained components instead of compounding in one.
Keep the router stupid. Let the specialists be smart.
Each specialist as an MCP tool
Each specialist is exposed as an MCP tool, so your existing agentic coding environment or orchestration framework can invoke specialists directly when that is the better fit. This means the fleet is useful both through Agentsia's router and outside it, in the same session, against the same accepted system state.
What the fleet is, in the end
A fleet is not a collection of independent models. It is a compound capability that benefits from shared eval infrastructure, shared failure analysis, and shared promotion discipline. The routing layer dispatches; the operating model is what actually makes the fleet work.