Build a research agent

Create a workflow

Click + beside Workflows in the sidebar. Right-click the new workflow, choose Rename, and name it People research.

The new workflow contains a Start block. Its built-in input field receives messages from the workflow test chat; you do not need to add a custom field for this tutorial.

Add and connect Agent

Open Toolbar and drag Agent into the workflow. Check that Start connects to Agent. If no connection appears, drag from Start's output handle to Agent's input handle.

Select Agent to open its fields in Editor. Choose a model that supports tools and structured output, and configure its credentials if prompted.

Configure Messages

In Messages, create these two messages. Use the role dropdown to choose System or User, and the plus button to add another message.

System

Research the person named in the user's message using the available search tools.
Find their location, profession, and education. Use Unknown for anything you
cannot verify. Include sources in your answer.

User

<start.input>

<start.input> inserts the message passed into Start. See References for using block outputs in another block.

Add search tools

In Agent's Tools, open Add tool... and add Exa and Linkup. Select Exa's search operation and configure the fields shown for each tool, including credentials where required.

The agent chooses which tools to call. A tool with a missing credential or required setting can fail when the agent calls it.

Test and inspect the run

Click Open chat beside the workflow's More actions menu. This opens the workflow test chat, which has an Outputs dropdown. The Chat tab in the panel is where you talk to Sim about building the workflow.

In Outputs, select Agent → content. Send:

Research the professional background of Tim Berners-Lee. Find his profession,
education, and a publicly documented location. Mark anything you cannot verify
as Unknown.

Expand Logs at the bottom of the builder and select Agent's row. Inspect its output, tool calls, and sources. Check that the results identify the intended person and support the returned facts.

Define structured output

In Agent's Response Format, paste this schema:

{
  "name": "person",
  "description": "A researched professional profile",
  "strict": true,
  "schema": {
    "type": "object",
    "properties": {
      "location": {
        "type": "string",
        "description": "A publicly documented location, or Unknown"
      },
      "profession": {
        "type": "string",
        "description": "The person's profession, or Unknown"
      },
      "education": {
        "type": "string",
        "description": "The person's education, or Unknown"
      }
    },
    "required": ["location", "profession", "education"],
    "additionalProperties": false
  }
}

You can also use the generation control beside Response Format to create a schema from a description. Check the generated schema before running it.

person names the schema. Its properties become outputs directly on Agent; it does not add a nested person field. Replace the System message with:

Research the person named in the user's message using the available search tools.
Return only the location, profession, and education fields defined by the response
format. Use Unknown for anything you cannot verify.

Keep the User message set to <start.input>.

Verify the structured result

Send the same prompt in the workflow test chat. In Outputs, select Agent's location, profession, and education fields. Open Agent's complete output in Logs and check their values before using them downstream.

When the model returns the schema's fields as valid JSON, they are available as <agent.location>, <agent.profession>, and <agent.education>. Structured output requests a shape; it does not verify the facts.