agent-graphs

agent-graphs is a LaunchDarkly skill for creating and managing agent graphs — directed graphs where nodes are AI configs (each an agent with its own model, prompt, and tools) connected by edges that define routing and handoff logic between them. It solves the problem of designing multi-agent workflows where configs route to one another, such as triage-to-specialist chains, intent-based routers, support escalation ladders, and pipeline processing stages. The skill guides the user to design the topology, create it with the right edges and handoffs, and verify the routing end to end.

The skill depends on the remotely hosted LaunchDarkly MCP server and its tools: required tools create, get, and list agent graphs; optional tools update or delete graphs and get or create the AI configs that serve as nodes. It documents the graph structure (edges with key, sourceConfig, targetConfig, and optional handoff data; a root config as entry point), a five-principle design philosophy (design before building, one agent one job, root config as router, meaningful handoff data, verify the full path), and a concrete workflow with a JSON example for a customer-support triage graph. It also covers edge cases (missing nodes, circular routing warnings, single-node graphs, edge updates) and anti-patterns to avoid.

It targets developers and platform teams building multi-agent systems on LaunchDarkly who want structured guidance for graph topology and routing. All actions are performed through official LaunchDarkly MCP tooling; the only destructive capability (delete-agent-graph) is an official, optional tool the skill advises using cautiously.