InferenceResearch

One gateway. Many models.

What is the right model for this job?

Every model call through one door

An agent platform that cannot see its own model calls cannot budget them, cannot audit them, and cannot stop one. The design thesis under study is one path for every call, rather than each agent holding a provider key.

That path would cost a hop and, if it holds, buy three things still under measurement: spend that can be refused before it is incurred rather than discovered on a bill, a record of what was asked and answered, and the ability to change model without changing an agent.

Choosing per role, not per company

A planner and a summariser do not want the same model, and pinning a whole company to one is how a platform ends up paying frontier prices for clerical work. Model selection belongs alongside the rest of a role’s declaration — what it may call, what it may spend — rather than in configuration somewhere else.

What we do not yet know is how much of the choice can be made automatically without the failure mode where a cheaper model quietly degrades a result nobody re-checks.

About this note
Status
Research
Focus
Model pick, price, and failover
Surface
Early / evolving
Papers
Findings land on /lab/papers when gated

Follow the work

Early Lab. Papers and findings land here when they clear the gate.