AI agents vs workflows: when to use which

An AI agent is given a goal, instructions and a set of tools, and a language model decides at run time which tools to call and in what order. It can deal with free text, with surprises and with requests nobody planned for. The price is that it can take a different path on different days, and every step costs a model call. What is an AI agent? explains the loop in more detail.

• The path

  Workflow

  Fixed, drawn in advance

  AI agent

  Chosen by the model as it goes

• Best for

  Workflow

  Repeatable processes

  AI agent

  Open-ended requests and messy input

• Cost per run

  Workflow

  Low; a model is called only where you add one

  AI agent

  A model call for every step

• Same input, same result?

  Workflow

  Yes

  AI agent

  Usually, not always

• How you test it

  Workflow

  Run each branch

  AI agent

  Try many example requests and read the tool calls

• When a step fails

  Workflow

  It is retried, then follows its error path

  AI agent

  The model reads the error and decides what to do

• You can write the steps down before you start.
• It must run on a timer, or when another system calls it.
• The result must be the same every time, for example for finance or compliance.
• It runs at high volume, where a model call per step would cost too much.
• The input is already structured, such as a form submission or a webhook from another system.

• The input is free text: an email, a chat message, a support ticket.
• The next step depends on what the last one found.
• Many different tools might apply, and choosing between them takes judgment.
• A person is talking to it and expects a sensible answer to an unplanned question.

Yes, and it is often the best design. Let the workflow own the trigger, the timing and the fixed steps, and call a model or an agent only for the part that needs judgment. The run stays predictable where it can be, and flexible where it must be.

A worked example: a website form sends each new enquiry to a workflow.

1. 1The workflow starts from a signed webhook when the form is submitted.
2. 2An “Ask AI” step reads the message and returns JSON with one field, the kind of enquiry, checked against a schema.
3. 3A switch step sends sales enquiries one way and everything else another.
4. 4For sales, an agent step researches the company and creates or updates the contact in the CRM. Its CRM writes wait for a person’s approval.
5. 5For everything else, an action step posts the enquiry to the support channel in Slack.

Only two of those steps call a model, and only one of them can choose its own path.

Workflows are drawn on a visual canvas in the console. These steps can run today:

• Action: one action of a connected tool, or one tool of an MCP server.
• HTTP request: a call to any HTTPS address, optionally with a key that is sent only to the hosts you list.
• Ask AI: one model call with no tools, answering in text or in JSON checked against a schema.
• Agent: runs one of your published agents and waits for it, including any approvals its tools need.
• Condition and switch: choose a branch.
• Set: store values for later steps.
• Delay: wait for a time or until a date, on a timer that survives restarts.
• End: finish the run as completed or failed, with an output.

A workflow starts from a webhook (signed with HMAC SHA-256 and a timestamp, or checked against a static token), from a schedule written as a cron line in your time zone, or by hand and through the API. Each run is a durable Temporal workflow that can last up to 30 days. Action and HTTP steps are retried three times by default with a growing wait between tries, and a step that still fails can follow its own error path.