Skip to content
Search is loading.

Select backends and models

Available now Each agent() call can select its backend and model.

Before you author model IDs or option values, use the live catalog. Do not infer these values from a model name.

Run this command to probe all configured backends:

All configured backends · catalog probe

npx -y @automatalabs/workflows config

Run this command to probe only Codex and Claude:

Codex and Claude · catalog probe

npx -y @automatalabs/workflows config codex claude

The command opens one no-prompt session for each backend. It does not send a model prompt.

The result lists the advertised model IDs, modes, and configuration options. A probe error for one backend does not stop the other probes.

If you accept the current default model, use the backend name:

Default Codex route · agent call

const result = await agent('Inspect the project.', {
label: 'inspect-project',
model: 'codex',
});

The model value selects the codex backend. The backend selects its current default model.

To select one advertised model for the call, use backend/model:

Explicit Claude route · agent call

const review = await agent('Review the implementation.', {
label: 'review-correctness',
model: 'claude/opus[1m]',
mode: 'plan',
configOptions: { effort: 'max' },
});

The backend prefix selects the ACP agent. The remaining value selects the model for that backend.

Copy option names from the catalog without changes. Each backend has its own option names.

These values came from the live @automatalabs/workflows@0.43.0 catalog on 2026-07-22.

Backend Model value Mode Configuration option
Codex codex/gpt-5.6-sol read-only or agent reasoning_effort: "max"
Claude claude/opus[1m] plan effort: "max"
pi pi/moonshotai/kimi-k3 No advertised Pi mode option thinkingLevel: "max"

Catalog values can change after a backend release. Before you copy these values into a new workflow, run config again.

Different calls can use different backends in one workflow. Each backend uses its own authentication source.

If a lower-cost route is sufficient for discovery, use it. Use a different route to review important work.

Structured output uses the same schema option across the bundled backends. AgentPrism validates the returned value against that schema.

After you select the routes, run the workflow validator. The validator checks the catalog for each routed backend.

Validate the workflow script.