Model Routing
Playbooks · LLM overview · ① Capability matrix →
Implementation guides for the LLM Blueprint. Plane ① on Agents may pin model_profile on the route row. The LLM gateway resolves that to a registry-approved endpoint. Principles: G.A.I.N LLM.
THE CLAIM
The model does not choose which model runs. Task-aware routing, abstention, and the capability matrix live on the gateway, not in the agent prompt.
Three playbooks
| # | Playbook | One-line purpose |
|---|---|---|
| ① | Capability matrix | Approved endpoints per profile, task, data class, and region |
| ② | Gateway task routing | Plan vs synthesize vs classify; failover, cost caps, abstain |
| ③ | Canary promotion | Eval-gated model swap, rollback, and change records |
Recommended path
- Capability matrix: version the approved pool before writing routing code
- Gateway task routing: filter, score, pick or abstain on every inference call
- Canary promotion: eval gates and rollback before a new endpoint becomes
stable
Bridge reading: G.A.I.N LLM · Inference handoff · Route contract reference · Eval Reasoning plane.
What this plane decides
| Decides | Does not decide |
|---|---|
| Which approved model endpoint for this call | Which workflow or manifest (Plane ①) |
| Cost, latency, region, data-class constraints | Which tool to propose (Plane ②) |
| Canary vs stable route for a task tier | PEP verdict on side effects |
Who should read what
| Role | Start with | Then |
|---|---|---|
| AI platform | Capability matrix, Gateway task routing | Canary promotion, G.A.I.N LLM |
| Domain squad | Capability matrix | Route model_profile in Route contract reference |
| Governance | Canary promotion | Model approval, eval sign-off |
| SRE / FinOps | Gateway task routing | Cost caps, SLOs, failover |