Graph

Every node you add to a Graph is a live LLM Agent and must be bound to an Agent preset.

Edges are not fixed pipeline routes. They represent permitted communication channels between Agents. Agents not connected by an edge cannot directly communicate or hand off work. Connected Agents can communicate and transfer state, but they do not have to do so. The actual behavior depends on Agent judgment and user instruction during runtime.

Edge descriptions are visible to Agents, helping them understand what the Agent on the other side is responsible for.

In the editor, the node inspector binds an Agent preset to a node and can append a node-specific prompt — so one preset can play different roles in different Graphs. The description you write in the edge inspector is injected at runtime as the handoff instruction, so write it in full: when to hand off, what to pass, and what the other side should produce. The canvas also supports validating the Graph, saving, and managing the blueprint ID, name, and description.

Each Agent preset bound to a node, together with the edges between nodes, defines a working crew — that reusable, editable template is its gene; see A-Genetic Engineering. And you don't have to drag every node by hand: the Atlas director or the built-in chat assistant can wire edges and add or remove nodes for you. When an Agent grows a Graph like this dynamically at runtime, that is Prolifera Engineering.