An agent is a published snapshot of a node graph. Every time a user sends a message in the chat UI, Raian executes the graph from a Chat Trigger and stops at a single Reply node to generate the assistant response.

Agents are different from bots (single prompt + tools) and workflows (event-driven automation). Agents are chat-triggered flows, built visually, and designed to respond in a chat surface.

Draft vs published

The agent builder edits a draft graph. Publishing creates a snapshot (versioned) that powers the public chat at /agents/[slug].

After publishing, you can keep iterating on the draft without affecting the public chat until you publish again.

Use Agent builder & runtime to learn how nodes work, how state persists, and how tools are configured.

MVP graph constraints

For the first release, agents follow these constraints:

Exactly one Chat Trigger node.
Exactly one Reply node, reachable from the trigger, and terminal (no outgoing edges).
No loops (the graph must be a DAG).