A step that decides what to do next
An agent node reasons, calls the tools you grant it, reads what comes back, and decides again — until it has an answer. The models are included, so there is no vendor API key to obtain before you can build one.
An agent node that calls real tools, on models included with your plan, with every reasoning step and tool result recorded.
What an agent can do
- Tools it can actually call — Draw an edge from the agent to a tool and it may choose to call it — once, many times, or never. The agent decides when; you decide what it is allowed to reach. Nothing is granted by default.
- Models included, no key to manage — Every workspace is issued its own gateway credential automatically, so an agent runs the moment you place it. There is no vendor account to open first, and no long-lived provider key sitting in your secrets for someone to leak.
- A trace you can read — Every cycle is recorded: what the agent was thinking, which tool it chose, the arguments it sent and the result it read back. When an agent does something surprising, the reason is in the record rather than in a guess.
- Bounded by design — An agent that loops cannot run up a bill. Each workspace has a monthly token allowance checked before the call is made, and a cycle limit that ends a turn that will not converge. Both are properties of the plan, not of the prompt.
Platform capabilities
- Included: Models, on every plan including the free one
- Per call: Tool arguments and results kept in the trace
- Checked first: Token allowance, before the model is called
Do I need my own OpenAI or Anthropic account?
No. Each workspace is issued a gateway credential when it is created, and your plan's monthly token allowance covers the usage. If you would rather run on your own provider account, you can supply your own key instead — but nothing is blocked until you do.
What stops an agent looping forever?
Two independent limits. A turn ends after a fixed number of reason-act-observe cycles whether or not it reached an answer, and the workspace's token allowance is checked before each model call, so an agent cannot spend past the plan it is on even if the cycle limit were removed.
Can I see why an agent did what it did?
Yes. Each cycle is stored with the reasoning, the tool the agent selected, the arguments it passed and the observation it read back. That is the difference between debugging an agent and re-running it hoping to see the problem again.