Skip to main content

Build an agent

An agent is the everyday shape: a chat assistant with a system prompt that can reach for tools when it needs them. This is the guide for going past a bare prompt to something that can actually do things.

  1. Start from the Agent pattern

    New flow → Agent. You get the core shape wired for you: chat in, a system prompt feeding a model, chat out. Write the system prompt — the instructions that define how it behaves and what it's for.

  2. Give it tools

    A prompt-only assistant can only talk. Add tools so it can act:

    • Web use — let it search the internet and read pages, for questions its training doesn't cover.
    • A knowledge base — ground it in your own content: search your documents, query your spreadsheets with SQL, or walk a knowledge graph. Pick the knowledge base and Workbench wires the retrieval for you.
    • Another flow, as a sub-tool — attach a flow you've built (a classifier, an extractor) so the agent can call it mid-conversation. Compose behaviors instead of cramming everything into one prompt.
    • A workspace integration — connect the agent to an external system your workspace has set up.

    The agent decides at run time which tool to reach for; you decide which are on the table.

  3. Parameterize the prompt

    Anywhere you write {{something}} in the system prompt, Workbench turns it into a typed input on the flow — so the same agent can be pointed at a different product, tone, or audience per run without editing it.

  4. Test, then publish

    Talk to it in the test pane until it behaves — including the cases where it should reach for a tool. Then publish a version so your integration pins a frozen snapshot.

  5. Put it to work

    Call it from your code (run a flow, chat shape), drop it into a channel (Slack), or hand out a share link. Same agent, wherever people are.

3 min read