01
Model choice
Match complexity, latency and cost to the model that should handle the request.
04 / Route
Route the Work
Make the next path explicit: use the fast model, the strong model, a tool, a person or no action.
Two routing layers
Next pathIncoming request
“Refactor auth for SAML and SCIM while keeping backward compatibility.”
Next path
strong model · tools
Model choiceList capabilities420 ms
Tool choiceAsk the router1.8 s
FallbackApply policypolicy
Two routing layers
Separate destination from permission
Choosing a tool is not the same as allowing it to run. Keep both decisions visible in the workflow.
02
Tool choice
Select the next approved tool or decide that no tool is needed.
03
Fallback
Send uncertain or high-impact requests to a stronger model or a person.
Operations
Optimize cost without hiding risk
A cheap route is useful only when the workflow knows when to step up.
Describe the routes
Describe the routes
Give every model or tool a short, honest capability description.
Ask focused questions
Ask focused questions
Use one question for complexity, destination and approval.
Include no-tool
Include no-tool
A router should be allowed to choose that no tool is appropriate.
Choose a route
Optimize cost without hiding risk
A cheap route is useful only when the workflow knows when to step up.
Choose a route
POST /v1/route{
"state": "Summarize this 80-page contract and flag renewal risks.",
"questions": {
"complexity": { "type": "score", "instructions": "How complex is this request?", "criteria": ["simple", "moderate", "hard"] },
"destination": { "type": "choice", "instructions": "Which route should handle it?", "criteria": { "fast": "short lookup", "strong": "complex reasoning", "tool": "needs a document tool" } }
}
}01
List capabilities
Describe the models, tools and handlers available in this loop.
02
Ask the router
Return a typed destination and the signals needed for the next gate.
03
Apply policy
Let deterministic code run, fallback or request confirmation.
FAQ
Questions about routing
Is this only for model selection?+
No. The same pattern can choose a model, handler, tool, queue or human review path.
Should routing and permissions be one question?+
Usually not. Ask where the request should go, then apply a separate deterministic permission or confirmation policy.
How do I handle low confidence?+
Choose a fallback such as a stronger model, a safe default or human review, and measure how often that path is used.
