Multi-Model Routing & Policy Orchestration
Executive Summary
This sub-scenario focuses on routing strategies for Planner/Orchestrator in multi-model, multi-tenant environments: automatically selecting primary and backup models based on task tags, cost, latency, and risk levels, supporting A/B testing and staged rollout, with quick rollback or degradation when anomalies occur, ensuring hit rate and experience.
Scope & Guardrails
- In Scope: Capability tags, policy configuration, decision API, fallback, safe mode, policy version management, rollback.
- Out of Scope: Provider onboarding (handled by Provider sub-scenario), cost billing details (handled by governance sub-scenario).
- Environment & Flags:
multi-model-router,routing-safe-mode; depends on Capability Graph, Feature Flag, Telemetry.
Participants & Responsibilities
| Scope | Repository | Layer | Responsibilities & Deliverables | Owners |
|---|---|---|---|---|
| planner-integration | powerx | integration | Decision API, trace, policy execution | Agent Platform Guild |
| policy-center | powerx | integration | Policy templates, versioning, audit | Agent Platform Guild |
| ops | powerx | ops | Monitor hit rate, trigger safe mode | Ops Reliability Center |
End-to-End Flow
- Planner inputs task tags → 2. Router selects primary/backup model based on policy weights and health scores → 3. Output decision/trace/cost estimation → 4. Monitor hit rate and fallback, roll back policy when necessary.
Key Interactions & Contracts
POST /internal/model-routing/route,POST /internal/model-routing/rollback,EVENT agent.routing.policy.updated.- Policy files:
backend/config/agents/routing/*.yaml,config/policies/model-routing.json.
Acceptance Criteria
- Hit rate ≥90%, fallback success rate ≥95%; policy change to effect <5 minutes; safe mode activatable within 1 minute.
Telemetry & Ops
- Metrics:
agent.routing.hit_rate,agent.routing.fallback_total,agent.routing.policy_rollback_duration. - Alerts: Hit rate drop, fallback failure, latency threshold exceeded.
References
docs/meta/scenarios/powerx/agent-and-automation/agent-model-platform/primary.mdbackend/config/agents/routing/*.yaml
