Models, providers, and routing

A model supplies capabilities such as text generation or image understanding. A provider supplies access to models through a configured service or local endpoint. A route selects the provider endpoint through which Holon will call a model.

This distinction matters when the same model is available through different services. The model name alone does not establish which credentials, endpoint, or service options the request will use.

Availability and capabilities

The model catalog describes models and their capabilities. An executable route also needs a usable provider configuration and, where required, credentials. Catalog presence does not prove that a route is available to this runtime.

Check capabilities for the intended operation. A conversation model may not support image input, image generation, or the Decision subsystem. A text task working successfully does not prove those other operations are configured. The model reference is the catalog entry point; configuration describes availability and credential inspection.

Global defaults and agent choices

The global default supplies the conversation route for agents without an override. An agent-level selection overrides that primary route for the agent without changing the global default.

For example, a documentation agent and a code-review agent can use different primary routes while sharing the same runtime. Changing one agent's model does not give it different permissions or transfer its work to a new identity.

Holon distinguishes a logical model reference from an executable route that also names an endpoint. Use the route required by the setting rather than assuming every model identifier is interchangeable.

Fallback and dedicated routes

The conversation provider chain starts with the agent override, or the global default when no override exists, followed by configured fallbacks. The global default is not silently added as a fallback after an agent override. Automatic fallback selection can use preferred routes from authenticated providers.

Fallback improves availability but can change the provider that receives a request. Inspect or disable it when a task requires a specific provider. It does not make providers equivalent in capabilities or output, and it does not guarantee recovery from every failure.

Image inspection, image generation, and advisory decisions have their own routing requirements. Configuring a conversation model is not a complete configuration for all tools. Dedicated vision and image-generation selections let those operations use suitable routes without replacing the conversation model.

Fast mode

On supported routes, Fast mode requests a faster service tier. It is separate from the model's thinking level and can consume more quota or cost more. Availability depends on the exact route; a compatible gateway does not automatically inherit the service's Fast capability.

Auto uses the route default. Standard can override a Fast default. Fallback routes use their own defaults, and an upstream service may serve a different tier from the requested one. Treat Fast as a request, not a latency guarantee; inspect requested and served tiers when diagnosing it.

Advisory decisions

The optional Decision subsystem can supply a bounded second opinion among specified choices. For example, an agent might ask which of several diagnostic checks to try first. The answer is evidence for the agent to consider.