Visual AI agent builder: the graph you draw is the graph that runs

A visual AI agent builder lets you define an agent as a diagram of steps instead of code alone: which node reads memory, which retrieves knowledge, which calls the model, which runs tools, how they loop and where the answer leaves. Each node has explicit settings, and edges define the order.

In AgentRow the diagram is executable. Validation runs the same rules in the dashboard and in the runtime, so a graph that passes validation is a graph the runtime accepts. There is no second, hidden version of the logic to drift away from what you see.

Every new agent starts with a working graph: input, memory read, knowledge retrieval, context builder, then the language model with a toolbox loop, and finalize. Memory write runs on a side branch after the answer so the reply is not delayed.

The LLM node has a conditional edge: if the model asked for tools, the run goes to the toolbox, which executes the calls and returns to the model for the next iteration, up to the agent's maximum iterations. Otherwise the run goes to finalize, the only node that streams text to the user. Internal branches never leak into the chat.

The Workflow tab has a graph editor: add nodes from the palette (input, context transform, memory read, knowledge retrieval, context builder, LLM, toolbox, router, validation, memory write, finalize), connect them, set per-node configuration and validate. Errors point at the node that needs attention.

Saving creates a draft version; publishing makes it active; earlier versions can be restored or archived. The builder's test chat can run drafts, and every test message creates a run you can open in Logs. Apps and end users only ever run the published version.