Define, govern, and monitor multi-agent systems as a single unit.

Real agent products rarely run on a single model and a single prompt. A typical workflow chains a router agent, one or more task agents, and a reviewer, with data flowing between them. Until now you had to track each agent on its own and reason about the system in your head.

Agent Graphs let you model the whole workflow inside AgentControl. You define each node as an existing agent configuration, draw the edges that describe how data flows between them, and version the graph the same way you version any other agent. You can target users, run experiments, and roll back the entire graph as one unit.

Agent Graph monitoring gives you a real-time view of how each node is behaving: latency at every hop, cost per run, quality scores from your judges, and which paths through the graph are getting taken. When a multi-agent product misbehaves, you see exactly where in the graph it broke.

Read the Agent Graphs documentation to get started.