Direct answer
A coding agent model router evaluates context before a session starts: which repository is involved, how risky the task is, what budget remains, which region applies, and whether a preferred account is near its limit. The result is a route the local CLI can apply.
When this matters
- A model with deeper reasoning is reserved for risky refactors and production incidents.
- A cheaper lane handles docs, tests, and simple UI changes.
- A team needs failover rules when a provider limit is close.
Operating steps
- Define model lanes in language engineers understand, such as fast patch, deep review, and guarded release.
- Connect each lane to allowed account references and CLIs.
- Set budget and limit thresholds that trigger fallback routes.
- Show the selected route before the session begins.
- Capture route changes in the audit log.
Common risks
- Fallback routing can accidentally downgrade safety if risk is not part of the decision.
- Routing only by cost can damage code quality for critical tasks.
- Routing only by model preference can ignore quota and regional constraints.
How AISwitchboard fits
AISwitchboard combines model routing with profile registry, local secret references, limit-aware switching, and rollback configuration.