Change the model of an agent in a workflow

Preset overrides vs custom mode, the media-type filter, External Agent routing, and what to check when the model list comes up empty.

Every agent step in a workflow picks its own model — entirely separate from the setup under Account settings → AI assistant. Within one workflow it is perfectly normal for one step to classify with a cheap model while another writes the report with a strong one.

Where to change it

Open the workflow editor and select that agent step on the canvas.
Open the agent configuration dialog.
Check the configuration mode at the top first: preset template or custom. The model lives in a different place in each mode.

Preset mode: overriding two fields is enough

Once a preset is chosen (universal, coder, researcher and so on), every field below it acts as an override — leave one empty and the preset's own value applies. Changing the model only takes two of them:

Provider — Pick the service provider. Change it and the model list is refetched for the new provider.
LLM Model — Pick from the list or type the model ID by hand. Left empty, the preset's default model stays in effect.

Change only the model, nothing else

Leave the system prompt, tool set and max iterations empty — empty means "follow the preset", so changing the model does not drag the preset's other settings along with it.

Custom mode: pick the routing first, then the model

Custom mode adds an extra layer, the External Agent routing block, where three options decide who executes this step:

Routing	Execution Method	How the model is chosen
Standard Agent	The platform's built-in engine, calling ordinary LLMs by provider and model.	Choose provider + model; the credential is resolved from the credential center.
Claude Code Agent	Runs the Claude Code command line directly.	Store the Claude subscription token in the credential center, then choose a model.
Codex Agent	Runs the Codex command line directly.	Store the Codex subscription (auth.json) in the credential center, then choose a model that account supports.

The three routings also treat the system prompt differently: the standard agent uses it as is; Claude Code appends it via --append-system-prompt; Codex has no separate system-prompt parameter, so the text is prepended to the task prompt and sent together.

Cannot find the model you want in the list?

Check the model type switch near the top of the dialog first — text, video or image:

With video or image selected, the list shows only models that support that media type, and typing a model ID by hand is disabled;
Only under the text type can you type any model ID directly.

So an empty model list usually has one of two causes: the model type is set to image or video, or you just changed provider and the list has not been refetched. Closing and reopening the dropdown triggers a reload.

How the credential is found

At run time it looks in the credential center for a credential of the same provider, in personal → team → system order. The workflow file never stores a plaintext key.
The credential center section inside the dialog lists which providers the current configuration references — just fill in what is missing there.
Older workflows may still carry a plaintext key inside their preset overrides; the page prompts you to migrate it into the credential center and then clean it up.

View full instructions for credential management

Custom model aliases

The dialog lets you register a custom model: give a provider + model ID pair an alias and it becomes selectable in the model dropdown. Handy for internal proxy gateways, self-hosted models, or new releases the official list has not picked up yet.

What it looks like in YAML

Every change on the canvas syncs into the YAML. If you write YAML directly, the model configuration sits on the agent definition:

yaml

agents:
  - name: summarizer
    preset: universal
    llm:
      provider: openrouter
      model: anthropic/claude-sonnet-4.5

  - name: classifier
    preset: lightweight
    llm:
      provider: deepseek
      model: deepseek-chat

provider takes the service identifier (openai, openrouter…), not the label you gave the credential. This is the single most common mistake here.

A few practical tips

Use a small model for deterministic steps like classification and extraction, and save the strong model for writing and reasoning — mixing models within one workflow is perfectly normal.
After switching models, do a dry-run or single-step execution to confirm the output format has not changed before putting it on a schedule.
For workflows shared across a team, reference team credentials — otherwise colleagues will not resolve a key when they run it.