Build your first agent

1. Create an agent. Go to Agents → New agent. Give it a name and a description.
2. Write instructions. The Instructions tab holds the system prompt: who the assistant is, what it does, how it should answer, what it must not do.
3. Pick a model. The Model tab selects the provider and model (see Models and providers), temperature, maximum output tokens and the maximum number of tool iterations per run.
4. Add tools (optional). The Tools tab attaches tools from your organisation's tool library (see Tools).
5. Attach knowledge (optional). The Knowledge tab selects knowledge bases the agent retrieves from (see Knowledge bases).
6. Set memory (optional). The Memory tab chooses whether the agent remembers nothing, remembers within a chat, remembers for the agent, or remembers per user (see Memory).
7. Review the workflow. The Workflow tab shows the graph. A new agent starts with the default workflow (Workflows: nodes, edges and the default graph), which already covers memory, knowledge, the model, the toolbox loop and the final answer. Edit it only when you need something different.
8. Test. Use the test chat to talk to the agent. Every test message creates a run you can open in the Logs tab.
9. Publish. The Publish tab publishes the current draft. Apps and end users only ever run published versions; the test chat can run drafts.

The agent builder tabs are: Basic info, Instructions, Model, Tools, Knowledge, Memory, Workflow, Logs, Guardrails, Approval, Schema, Runtime and Publish.