# Agents (/agents)





Build a workflow agent by connecting one or more **Agent** blocks to the steps it needs. Each Agent block sends messages to a model, lets it call configured tools, and returns a result for later blocks.

## Build a workflow agent [#build-a-workflow-agent]

A simple agent reads an input, scores it, and returns the score:

<WorkflowPreview workflow="BUILD_AGENT_WORKFLOW" height="200" highlightBlock="agent" />

1. Add an [Agent block](/workflows/blocks/agent) after a [Start trigger](/workflows/triggers/start).
2. Select a **Model**. In **Messages**, use a system message for the task and rules, and a user message for the input, such as `<start.input>`.
3. Add the tools or context the task needs, then run the workflow and inspect the result.

Without **Response Format**, later blocks read the answer as `<agent.content>`. Model, token, tool-call, and cost details are also available. The example requests a structured score, so its Response block reads `<agent.score>` after successful JSON parsing. See the [Agent reference](/workflows/blocks/agent) for model selection, settings, and outputs.

## Add capabilities [#add-capabilities]

| Need | Configure |
| - | - |
| Take an action | Attach integration tools, [custom tools](/agents/custom-tools), [MCP tools](/agents/mcp), or another workflow as a tool. The model decides which available tools to call. |
| Follow a reusable procedure | Select [skills](/agents/skills), such as a lead-scoring rubric. The agent sees their names and descriptions, then loads instructions when they apply. |
| Answer from your documents | Add the [Knowledge](/integrations/knowledge) tool and select the knowledge base to search. |
| Read a file | Use the Agent block's **Files** input with a model that supports the file type. |
| Remember earlier runs | Configure [memory](/workflows/blocks/agent#memory) and a conversation ID. The memory mode controls which history reaches the model. |
| Return fields for later blocks | Set **Response Format** to request a JSON shape, then reference successful structured fields directly, such as `<agent.tier>`. Provider support and malformed-output behavior vary. |

Tools and skills need a compatible model and the required workspace, credential, and organization access. Adding a skill supplies instructions; adding a tool gives the model an action it can call.

## Choose where an action runs [#choose-where-an-action-runs]

Use a separate workflow block when an action must be attempted whenever that path runs. Attach it as an Agent tool when the model should decide whether it is needed. For example, a lead-scoring agent might search for missing company information, while a Google Sheets block records every score that reaches it.

The same integration can serve either role. See [Choosing what to use](/agents/choosing) for the differences among fixed steps, Agent tools, custom tools, MCP connections, workflows as tools, and skills.

## Related docs [#related-docs]

* [Agent block reference](/workflows/blocks/agent)
* [Choosing what to use](/agents/choosing)
* [How blocks pass data](/workflows/data-flow)
* [Chat](/chat)