What an agent is (and how it differs from an automation)

An agent is AI that acts on your records: it reads them, reasons, and changes them. An automation is the deterministic engine underneath, a trigger followed by steps that run the same way every time. The two meet when one of those steps hands control to an AI model, and that is where an automation becomes agentic.

Requirements

None to read this page. To add an AI step to an automation: a role that can edit automations, and an AI provider connected to the business; see Connect an AI provider.

Two words that are not synonyms

An automation (elsewhere: workflow) is predictable. When an entry moves to Placed, it emails the client and updates the recruiter's count. Given the same input, it does the same thing. Agents and automations from first principles builds that model.

An agent (elsewhere: AI assistant, AI worker) is not predictable in that sense, and is not meant to be. It reads the situation and decides what to write. The test that separates an agent from AI that only talks: does it change the record, or only describe it?

Two ways an agent reaches your records

It is one concept with two entry points.

• From inside Sonta, on an event. An automation reaches an AI step and hands the work to a model. The trigger decides when; the model decides what. A new inbound lead arrives, and the AI step researches the company and fills in its industry and size.
• From outside, over MCP. An AI assistant you connect, such as Claude, operates the business directly: it reads entries, creates them, and configures collections and automations, with the permissions you give it. How an agent connects and what it can reach covers this path.

You can also give a task to the agent inside the product directly, without building an automation first.

What an AI step does

An AI step sends a prompt to a model from a connected provider and puts the answer to work. It produces one of three things:

• Text: an answer that later steps can use, for example as the body of an email.
• An email: a subject and a body, written for a send-email step that follows.
• Field values: the model fills in fields you choose on an entry. It can set a reference by naming the record it means; if that name matches several records, the step fails rather than guess.

The step can also be allowed to search the web, with a cap on how many searches it may make. How to set one up is in Add an AI step to an automation.

Why keep both

Deterministic steps are the right tool when the outcome must be the same every time: assigning an owner, moving a stage, sending a confirmation. An AI step earns its place where judgment is needed: summarizing an email thread, drafting a reply in context, classifying a lead that does not fit a rule. Most useful automations mix the two, with the rules around the AI step deciding when it runs and what happens with its answer.

Notes and limitations

• An AI step needs a connected AI provider. Calls go to that provider and may cost money there.
• Test runs simulate the AI step and show the prompt that would be sent, without calling the provider.
• Some fields cannot be filled by AI. The step reports which ones it could not fill.
• A long answer can be cut off at the output limit; the step marks it as truncated.

Related

To build the automation around an AI step, see Add an AI step to an automation, Create and activate an automation, and Automation triggers.