Model providers
A model provider serves inference requests. The harness’s provider adapter translates its messages, request options, and expected response into the provider’s API format. A model is the specific model selected for the request. A provider account supplies the credentials and access needed to use it.
Provider and account resolution
Section titled “Provider and account resolution”What a provider adapter does
Section titled “What a provider adapter does”An adapter turns a harness request into the provider’s format and returns the result, tool requests, and usage in a form the runtime understands. Model capabilities depend on the chosen model and endpoint. The workspace or authorized execution actor supplies an account connection under its own policy.
| Check | Why it matters |
|---|---|
| Provider and model are supported together | A valid account may lack the requested model or API capability |
| Actor can use the selected connection | Team access does not grant every member a private provider account |
| Run has capacity and budget | An otherwise valid binding may still be blocked at dispatch |
Example selection
Section titled “Example selection”This example illustrates the selection decision a runtime can record. Use provider/model IDs and account references from the workspace configuration.
Step: draft meeting summaryNeeds: text input and schema-constrained outputBinding: workspace-approved text model and actor-authorized accountFailure: if schema-constrained output is unavailable, hold the step for setupConfiguration changes
Section titled “Configuration changes”Resolve provider bindings again before a new attempt. Record the binding used for the actual attempt so cost and behavior can be explained. If the configured account is revoked while work is waiting, recheck it before use. An alternate provider may have different tool or output support and should be checked against the same requirements.
A workspace can also bind to a model server it operates on the same machine or a private network. The adapter still needs to verify the model and API capabilities for the step. Local AI
See models, resource binding and keys, and the proposed binding rules.