Skip to main content
Gradient is the operating system for conversational agents. You build an agent as a graph of prompt nodes, ship it over phone and chat through one shared engine, and improve it with testing and evals that run in the same loop, not a separate tool bolted on afterward. Voice and chat are the same agent. The graph, prompts, tools, and published versions are identical across every channel; only the transport differs.

​
The loop

  • Build. Author an agent as a topology of single-prompt nodes. Each node binds its own model, system prompt, and tools, and hands control to other nodes as the conversation moves.
  • Publish. Freeze the draft into an immutable, numbered version. Live traffic only ever resolves to a version you shipped.
  • Run. One engine runs that graph over dashboard chat, the public API, and phone calls.
  • Evaluate. Redteam the real topology against a forked copy of your data, then grade both test and production conversations against one rubric.

​
Start here

Quickstart

Build, publish, and call your first agent end to end.

Core concepts

The vocabulary: nodes, topology, versions, datasets, rubrics.

Build agents

Author nodes, wire handoffs, and attach tools in the builder.

Write rubrics

Define what correct means and grade every conversation against it.