AI agents and models
An AI step in a flow is not configured from scratch every time. It names an agent, and the agent supplies the standing part of the configuration—who the agent is, how it should behave, what tools it may reach for, and which model its work runs on. The step supplies only the task.
That split is the point. Ten steps across five automations can share one agent, and changing that agent's instructions changes all ten. It also keeps each step's own form short: a prompt, and whatever that particular step needs on top.
What an agent is
An agent is a reusable record that you own and edit, kept on the Agents page. It holds:
| Field | What it does |
|---|---|
| Name | What the agent is called wherever it is offered. |
| Description | A line shown under the name on the catalog card. |
| Category | The heading the agent is filed under, and one of the filters on the catalog. Pick a heading another agent already uses, or type a new one. |
| Model | The account and model this agent's work runs on. See How a model is chosen. |
| Instructions | Standing instructions applied wherever this agent runs, on top of which each step writes its own prompt. |
| Tools | The MCP servers every step using this agent may draw tools from. |
A step's reference to an agent is live, not a copy. Edit the agent's instructions and every step using it changes with it, in the next run, with nothing to republish. This is a feature when you are refining how an agent behaves and a hazard when you are not: an agent used in ten places is edited in ten places at once.
The Agents page presents them as a grid of cards showing each agent's name, description, the opening of its instructions, and badges for its model and category. Narrow the grid with the search box and the category filter, and show archived agents when you need one back.
Presets are starting points, not links
Creating an agent from scratch gives you an empty record to fill in. New from Preset instead opens a gallery of ready-made starting points, each with a name, a description, a category, and a note on the kind of model its work suits. Choosing one creates an agent already filled in.
A preset is a seed, not a link. The agent it creates is yours from that moment on: edit it however you like, and later changes to the preset never reach it. This is the opposite of the step-to-agent relationship above, and the distinction is worth holding on to—steps track their agent, agents do not track their preset.
Presets deliberately leave the Model unset, because which account serves the work is yours to decide.
How a step departs from its agent
A step can differ from its agent in three ways, and the asymmetry between them is deliberate:
- Model Override—run this one step on a different account and model. Leave it unset to inherit the agent's.
- Ignore Agent Instructions—run this step without the agent's standing instructions. Your prompt still applies.
- Extra tools—a step may name additional MCP servers beyond its agent's.
A step can only add tools, never remove them. An agent's instructions are written assuming the tools it declares are available; a step that could take one away would break those instructions silently, with the agent simply failing to do a part of its job and saying nothing about why. Adding is safe, subtracting is not, so only adding is offered.
Whenever a step departs from its agent, a note at the top of the step's form says how—the model it will run on and which tier chose it, whether the agent's instructions are suppressed, and how many tools the step added. A step running its agent exactly as configured says nothing, because there is nothing to report.
How a model is chosen
Every AI step runs on some account with some AI provider. Which one is decided in three tiers, most specific first:
- The step's Model Override, on the step's own form.
- The agent's Model, from the agent the step names.
- The Default Model, on the Nodes tab of the Settings page.
The first tier that is filled in wins, and a tier left empty falls through to the next. With a Default Model set and nothing else configured, every step runs on it; an agent with a Model overrides that for its own steps; a step's own override beats both.
A tier that is filled in but cannot be resolved fails the run. It does not quietly fall through. If a step names an account that has since been deleted, or one that can no longer be reached, or an account whose credentials are incomplete, the step fails and says so. Falling through would put your content on an account you never picked and the cost on a bill you never agreed to—not something to do quietly.
With nothing configured at any tier, a run fails saying that no model is configured for the step, its agent, or as a default.
You do not have to work the tiers out by hand. Where a step's model differs from its agent's, the note at the top of the form names the model, the account serving it, and which tier decided. When the configuration cannot be resolved at all, that note becomes a red warning carrying the same message the run would have failed with—so you find out while editing rather than at three in the morning.
The three fields of a model choice
Wherever you choose a model—on a step, on an agent, or in the settings—you are filling in the same three fields:
| Field | What it does |
|---|---|
| Provider Account | The provider account whose credentials serve the model, and whose bill the tokens land on. |
| Model | The provider's own identifier for the model. Pick one the account offers, or type an identifier it does not list. |
| Reasoning Effort | How hard a reasoning model should think before answering: Minimal, Low, Medium, or High. Leave it blank to let the model decide. |
Model and Reasoning Effort stay hidden until you have picked a Provider Account. The list of models is generated from the account, so before one is chosen there is nothing to offer—the fields appear as soon as you make that first choice.
Set Reasoning Effort only where you know the model supports it; most models are not reasoning models and reject it outright. It is a per-step decision by design, so a cheap yes-or-no judgment and an expensive drafting step can run at different depths on the same account.
What it costs
Whichever provider account serves a step, the work is charged by that provider on their own terms. What the platform reports is token usage, and the approximate cost it implies, so the figures on the dashboard and in a run's tooltip are a guide to consumption rather than a bill. See Monitoring and history.
Where to go next
- For the accounts themselves—what each provider needs, and how to add one—see Integrations.
- For the steps that run an agent, see the AI Agent action in Actions and the AI Judgment condition in Conditions.
- For pausing an agent step to ask a person something, see Human in the loop.