Model Sets
Model Sets define the routing logic that picks a Model Profile for every agent invocation. They let you:
- Control cost by steering routine or high‑volume tasks to cheaper profiles.
- Specialize by sending domain‑tagged work to tuned or engineered profiles.
- Experiment safely by adding narrow rules that direct a slice of traffic to trial profiles.
Quick Reference
Model Set
| Field | Purpose | Guidance |
|---|---|---|
| Name | Identifier for routing rules | Human readable name for the model set. |
| Description | Purpose & routing intent | Document this Model Set. Good information to cover includes cost strategy & specialization goals. Make sure this is updated as rules evolve. |
| Default Profile | Fallback model profile | Choose a model profile handling the majority use cases efficiently. |
Selection Rules
| Field | Purpose | Guidance |
|---|---|---|
| Model Complexity | Route by agent complexity | Map low complexity tasks to cheaper models. |
| Tag Matching | Specialized profile selection | Use tags (e.g., domain:code) to route to tuned profiles. |
| Model Profile | Selected model profile | Choose the model profile to use when an agent matches the selected complexity and tag criteria. |
Core Concepts
Every set declares:
- Default Profile – the fallback when no rule matches.
- Rules – evaluated top‑down; first match wins. Order is significant, so put the most specific conditions first.
Rules may reference profiles from different providers. Mixing vendors (e.g., Gemini for summarization, Anthropic for code reasoning) lets you optimize for capability, latency, or price within one set.
Design Guidance
- Order rules from most specific to most general to avoid unintended shadowing.
- Keep the tag taxonomy lean; merge overlapping tags that don’t change routing behavior.
- Use the description field to capture intent, version, or experiment notes for later audit.
- Introduce experimental profiles behind clearly scoped tags (e.g., “exp-summarize”) so rollback is a single delete.
- Periodically review unmatched traffic; a growing fallback rate can signal missing tags or stale rules.
Related
- Model Providers & Profiles
- Agents (complexity & tags)
- Meshes