Build your first agent

Agents are LLMs that the platform routes to with a goal, a tool catalog, and access policy. They decide which tools to call and in what order — the opposite of a workflow's pre-wired graph. See Workflows vs agents and Agents for the longer version.

This walkthrough creates a "Deal prep" agent that, given an opportunity, looks up the account and the most recent transcripts and produces a short pre-meeting brief.

1. Open the Agent Builder

Navigate to Automation → Agents and click New agent.

2. Configure the agent

In the agent's configuration tab, fill in:

• Name: Deal prep
• Description: Given an opportunity, summarize the latest context for a pre-meeting brief.
• Inputs: declare opportunity_id so the agent knows what to expect from callers.

3. Pick a model

In the model section, choose a strong reasoning model. The platform's model list lives behind Settings → AI Models and follows the per-org default unless you override it here.

4. Give the agent tools

In the Tools tab, add:

• Get Record — so the agent can fetch the opportunity, the linked account, and any contacts.
• Get transcripts — so the agent can pull recent meeting transcripts for the account.
• A workflow tagged "Use as Tool" — for example the workflow you built in Build your first workflow, if you want the agent to be able to research the account on demand.

The Tools tab includes a built-in help drawer covering tools, agents, workflow tools, core tools, and how to choose the smallest useful tool set. Picking the smallest set that gets the job done is the single most important agent-design decision — extra tools cost tokens and make the agent more likely to wander.

5. Write the prompts

Agents have:

• A system prompt — the durable identity, role, and rules.
• An instructions prompt — the task for this run, which can interpolate workflow variables ({{ inputs.opportunity_id }}).

Keep the system prompt short and stable. Use the instructions prompt for the per-run "what to do this time" content.

6. Test the agent

Use the Test tab to run the agent with sample inputs. The trace view shows every tool call, every model output, and every retry — this is the fastest way to find prompts that don't work or tools the agent never calls.

7. Where the agent runs

The same agent configuration can run in three places:

• Genie surface — embedded in pages, typically used for product help and natural-language UI.
• Slack / Teams — when configured under Settings → Integrations.
• Workflows — via the Agent Task, which executes the agent as one step in a larger workflow.

8. Publish and use it

Once you're happy with the trace output, Publish the agent. From there:

• Run it from a record page (Records → Opportunities → Run agent).
• Add it to a workflow with the Agent Task so it becomes a step in a larger automation.
• Pin it to a Genie surface so users can call it by name in chat.

Next steps

• Browse the Concepts section for deeper background on triggers, tasks, and variables.
• Build a more realistic agent with multiple tools — see Build a multi-tool agent.
• Let your agent call a controlled workflow action — see Use a workflow as a tool inside an agent.
• Look at the Task reference to see the full set of building blocks workflows can call.