Models
A model performs the inference a harness needs to carry out an agent’s work. A model may interpret context, produce structured output, or propose a tool call. The harness supplies the context and controls the actual invocation of tools. A model provider serves the selected model through an authorized account.
Match the model to the step
Section titled “Match the model to the step”Inputs, outputs, and constraints
Section titled “Inputs, outputs, and constraints”Select for the actual task. A document extraction step may need vision input and structured output. A writer can use a text model with a smaller input. A spoken interaction can use a realtime audio model or a combination of speech and text services. Tool calling, input size, output validation, and cost are part of selection.
| Step | Required input | Required output |
|---|---|---|
| Summarize accepted calendar data | Structured data represented in text | A summary validated against named fields |
| Inspect a shared tab | Masked image and the visible control list | A browser action that passes validation |
| Give a spoken answer | Audio or a transcript | Audio, with tool calls checked by the same service |
Example capability check
Section titled “Example capability check”Illustrative decision logic for a connector-backed reasoning step.
if model cannot accept the requested input: hold for setupif model cannot return the required output: hold for setupotherwise run within the step budget and validate the returned fieldsRecord and verify the result
Section titled “Record and verify the result”Record the binding used
Section titled “Record the binding used”Model output is variable. The harness can check output structure, limits, and tool permissions. A specialist can still need to verify a claim against its source. Capture the selected model and actual usage for an attempt. ### Evaluate a replacement model
Test a replacement model with the same inputs, tools, and evals before using it for the same playbook. Check both output quality and failure behavior. For example, when a source is missing, a replacement summarizer should request the evidence instead of inventing a citation.
See model providers, evals, and Coach’s vision model.