Build your first agent

This guide walks through creating a single working agent end-to-end. It assumes you already have a project — see Quick Start if not.

1. Connect a provider first

An agent needs a provider configuration to call. If you haven't connected one yet, do that first — see Connect a provider.

2. Create the agent

Go to Project → Agents → New Agent:

POST /api/v1/projects/{id}/agents
{
  "name": "Support Assistant",
  "llm_provider_config_id": "<provider_config_id>",
  "temperature": 0.3,
  "max_tokens": 1024
}

3. Write the system prompt

Keep the first version focused and specific. You can iterate — every edit to the system prompt creates a new agent version, so nothing is lost.

You are a support assistant for UnderOcean. Answer questions about the
platform's features precisely. If you don't know the answer, say so rather
than guessing.

4. (Optional) Attach tools

If the agent needs external capabilities (e.g. searching a ticket system), assign MCP tools under the agent's configuration — see MCP.

5. (Optional) Attach a knowledge base

If the agent should answer from your own documents, attach a knowledge base — see Set up a knowledge base.

6. Run it

Open the agent's chat view and send a message. You'll see the response stream token-by-token, with an inference-steps panel showing any tool calls along the way.

Build your first flow to combine multiple agents into a single orchestrated workflow