Getting started

New workspace, nothing in it yet? This is the fastest path from empty to a working agent. In about 15 minutes you will add a model, create an agent, and get a real answer back in the Agent Playground.

Already have agents in your workspace? Jump to the Quick start — it covers running an existing agent instead of building one.

You do this inside a project (team workspace). If your sidebar only shows Home and Profile, open a project first (Teams and access).

Add a model

An agent needs a model to run on. Models are registered once per workspace and then reused everywhere.

1. In the workspace sidebar, open Models & tools → AI Models, then click Add model.
2. Give it a name your team will recognise (e.g. gpt-4o-eu).
3. Pick a provider (OpenAI, Azure, Anthropic, Google, AWS, Hugging Face, or Cohere) and enter the provider's model ID.
4. Choose an authentication strategy (API key, IAM, service account, or OAuth) and fill in the credentials it needs.
5. Leave the model type as chat and make sure it is active, then save.
6. Back on the list, use the row action Test model to confirm the credentials work.

Tip: Credentials are stored securely and never returned to the browser. If Test model fails, re-check the model ID and auth values with your provider.

Create an agent

With a model in place, build the agent. The New agent wizard walks you through it step by step.

1. In the sidebar, open Agents → Agents, then click New.
2. Identity — set a name (lowercase letters, numbers, and hyphens), an optional description, and choose the visibility and scope (personal vs team).
3. Model — select the model you just added. You can also create one inline here if you skipped ahead.
4. Prompt — pick an existing prompt or create one that tells the agent how to behave.
5. Skills — optionally attach skills the agent can use.
6. Review — check the summary and Create.

You now have a versioned agent. Editing it later creates a new version, so you can always compare and roll back (Safe rollouts).

Test in the playground

The Agent Playground runs your agent for real — no deployment required.

1. In the sidebar, open Agents → Agent Playground.
2. In the panel, select your agent, its version, and an environment (start with staging if your workspace has one).
3. Send a real prompt — a question your team actually gets, not "hello".
4. Read the response, then expand the trace / timeline to see the model call, any tools that ran, and the agent's reasoning.

If the answer is thin or wrong, that inspection tells you what to fix next.

If something blocks you

Next steps

• Prompts and resources — sharpen instructions and add knowledge.
• MCP servers and tools — give the agent real actions.
• Testing and release — validate before you ship.
• Safe rollouts — promote to production with confidence.

← Tour of the console · Quick start →