Agent graphs bring control and visibility to multi-agent AI workflows

Multi-agent systems offer the ability to distribute complex tasks among specialized agents for better efficiency, but managing performance across these agents can be challenging as the system scales. The introduction of agent graphs in AI Configs facilitates multi-agent workflow management by integrating it into the same control plane used for releases, experiments, and guardrails. In an agent graph, nodes represent agent-based AI Configs, and edges define how outputs are transferred between agents, enabling coordination, execution order, and reuse without duplication. The LaunchDarkly AI SDK evaluates each agent using standard targeting rules, while the application manages execution, allowing flexibility with existing frameworks. This structure allows workflow visibility and changes without altering application code. The GA release enhances this with agent graph monitoring, presenting performance metrics like latency and tool calls directly on the graph visualization. This allows for easy identification of bottlenecks or quality issues, which can be addressed within LaunchDarkly by adjusting node configurations without new deployments. The AI Configs control plane supports gradual rollouts and fallback variations to mitigate risks, ensuring rapid adaptation to changes. Agent graphs are currently available in AI Configs with Python support, and Node.js support is forthcoming.