Model orchestration for agents: runtime, routing, and evaluation

A reliable agent product needs more than a model and a tool loop. Runtime, routing, and evaluation solve different operational problems and should remain separate in the architecture.

Runtime operates the workflow

The runtime coordinates tasks, tools, context, retries, review points, and delivery. In Machina, Studio is the workspace for managing projects, agents, workflows, activity, and usage. SportsClaw is an optional open-source developer engine for applications that want an inspectable runtime component close to their interface; it is not the Machina platform.

Routing serves model requests

The Machina AI Router sends requests through configured providers and profiles according to the task and operating constraints. Routing can keep a product from hard-coding every workflow to one model. It does not create a new public OpenAI-compatible endpoint, and it is not an automatic optimizer.

Evaluation compares configurations

Arena runs evaluations and compares agents, workflows, or model configurations against relevant tasks and criteria. A useful evaluation names the dataset, expected behavior, scoring method, and failure cases. A generic benchmark can inform model selection, but it does not replace testing the real workflow.

Keep the decision human

Evaluation evidence can support a change in configuration. The current capability should not be described as a fully automatic registry-to-promotion loop or continuous self-improvement system. Owners still decide what changes, under which controls, and when.

Explore Studio, Arena, and the Machina AI Router as separate roles in one product family.