From nothing to a working agent
The whole path, in about ten minutes.
On this page
Four steps. Nothing here needs a developer, and none of it is irreversible.
1. Connect something
Go to Connections and add whatever the agent should work with — a database, an API with an OpenAPI document, an MCP server, or the demo dataset if you just want to see the shape of it.
AgentOS reads what that software can do and turns each capability into a tool. You do not write the tools; you approve which ones exist.
2. Make an agent
Agents → New. Give it instructions in plain language — what it is for, what it should refuse, who it is talking to — and tick the tools it may use.
Fewer tools is better. An agent with four relevant tools behaves far more predictably than one with forty, and it is easier to reason about when something goes wrong.
3. Test it before deploying it
Every agent has a test panel. Ask it something real and watch the timeline: you will see each tool it proposed, the arguments, and what came back.
This is the step people skip and then regret. A minute here saves an afternoon.
4. Deploy, then watch the first few runs
Deploying makes a version live. The previous version is kept, and you can roll back to it without redoing anything.
Watch Executions for the first day. Most surprises show up in the first ten real requests, and almost all of them are instruction problems rather than platform problems.