Building an Agent

The Agent builder walks you through everything an agent needs, section by section, with a live preview beside you the whole way.

Opening the Agent builder

1. From the sidebar, go to Agents.
2. Click New, then Build an agent.
3. The builder opens with its sections on the left and a live preview on the right.

[Screenshot: Agent builder showing the section list and live preview]

The Agent builder sections

The builder is organized into sections. You can move between them freely - nothing is locked - and the Continue to ... button at the bottom takes you to the next one. Your progress saves automatically.

1. Identity

Who this Agent is and how it sounds.

• Name, Role, and Summary - a clear name (e.g., "Acme Onboarding Concierge"), its job title, and a sentence or two on what it does. You can also change its avatar.
• Persona - traits that describe the character. Who is this Agent? What do they sound like?
• Vibe - four sliders (Formal to Casual, Terse to Chatty, Serious to Playful, Predictable to Creative) that shape the writing style and, when voice is on, the speaking style.
• Conversational voice - on by default. Pick a voice from Your voices or the Library. See Voice Mode for the full guide.

Tip - Be specific. "Helps customers" is too broad. "Answers pre-sale questions about Acme Pro, qualifies leads, and books demos" is much better.

[Screenshot: Identity section with name, persona, and vibe]

2. Goals

List what success looks like for this Agent, one goal at a time. Everything else (knowledge, tools, guardrails) flows from these.

3. Knowledge

Choose which Knowledgebases this Agent can read from at runtime.

• Create a new Knowledgebase or link an existing one. Each one holds documents, web pages, or videos.
• The Agent only searches the Knowledgebases you attach here, not everything in your workspace. This is how you keep an HR Agent focused on HR docs and a Sales Agent focused on the catalog.
• Linked Knowledgebases are shared: you can attach the same one to multiple Agents, and edits sync everywhere.

See How Search Works for the mental model behind retrieval.

4. Model

Choose how much reasoning the Agent uses: Smarter (deeper reasoning, slower), Balanced (one steady model on every turn), or Faster (snappier replies; recommended). You can also pick a specific model from the model list.

The Model section also holds the Built-in tools, each showing its credit cost per use:

• Web Search and Web Scrape - on by default for new Agents.
• Deep Research - multi-step research that compiles a cited digest.
• Image generation & editing - create new images and edit ones people upload, with a choice of image model.

You can change the model later without losing any of the rest of your setup. See LLM Providers to bring your own provider keys.

5. Integrations

Give your Agent apps it can act in during a conversation - for example, look up a customer or draft an email.

• Connect apps through Composio. This section only appears when external services are enabled for your workspace.
• Under Actions, pick which specific actions to expose for each app - e.g., from Gmail you might allow "draft" but not "send."

Tip - Start with one or two apps. Too many tools makes the agent less reliable, not more.

6. Orchestration

Link Flows the Agent can call for help. Linked Flows become tools, and the Agent hands off to one when it's a better fit for the request. You can link an existing Flow or create a blank one here.

7. Guardrails

Define what the Agent should never do, grouped by area: Identity, Goals (out of scope), Knowledge, Integrations, and Orchestration. Examples: "Never discuss pricing - always defer to sales." "Claim to be a human."

The Clarify panel at the bottom can Run clarification: the Agent reviews your goals and guardrails, asks up to four questions about vague spots, and proposes approval rules for each connected action - by default, read actions run automatically, write actions ask the first time, and destructive actions always ask.

8. About

What end users see before they chat: an introduction page, up to eight Conversation starters shown as tappable chips, and an optional chat disclaimer under the chat input. Changes here reach end users on the next publish.

Publish and share

Once everything looks right, publish your Agent.

• Use the live preview on the right to chat with it as you go. The preview reflects your latest changes.
• When you're happy, click Publish and Share in the header (it becomes Publish after the first release, with a separate Share button). The Agent becomes available to share or embed (see Deploying for sharing options).

Live preview

Every section has a live preview pane on the right. As you tweak the persona or attach knowledge, send a test message - the response updates against your latest configuration. Use this to:

• Sanity-check guardrails ("Is it refusing the right things?")
• Verify retrieval ("Did it find the right doc?")
• Try voice ("Does the picked voice match the vibe?")

Tips

• Iterate, don't perfect. Ship a first version, watch real conversations in Execution History, then tune persona and guardrails based on what people actually ask.
• Keep the purpose narrow. Narrow Agents beat broad ones. If your Agent is trying to do five jobs, split the work into focused Agents and Flows.
• Don't over-attach knowledge. More sources mean noisier retrieval. Attach only what's relevant.

Next steps