Multi-agent workflow graphs

Multi-agent software splits a job across several artificial intelligence agents, and LaunchDarkly says its Agent Graphs feature is now generally available to map those handoffs visually. (launchdarkly.com) In LaunchDarkly’s model, each node is an agent-based AI Config and each edge defines how output moves from one agent to another. The company says the graph lives in its control plane while the customer’s application still handles execution. (launchdarkly.com 1) (launchdarkly.com 2) That split is the selling point: workflow structure can be changed outside application code. LaunchDarkly’s March 20, 2026 tutorial says topology updates made in the user interface are picked up by traversal code on the next request, without a redeploy. (launchdarkly.com) The company says the general-availability release adds monitoring directly on the graph, including per-node latency, invocation counts, and tool calls. Instead of matching logs across separate systems, teams can see performance data on the workflow diagram itself. (launchdarkly.com) LaunchDarkly also says the software does not run the agents for customers. Its documentation says the platform resolves graph structure and evaluates agent configuration, but state handling, traversal decisions, model calls, and termination logic all stay outside LaunchDarkly. (launchdarkly.com) That boundary matters because the hardest production problems in agent systems often sit in execution history, not just layout. Microsoft Research wrote in March 2026 that multi-agent failures are long, stochastic, and can bury the first unrecoverable error step deep inside a trajectory. (microsoft.com) Researchers are also separating saved state from provenance, the record of how an agent reached a decision. A March 2026 paper on reasoning provenance argued that execution state alone cannot reliably reconstruct the reasoning path after the fact. (arxiv.org) LaunchDarkly’s own tutorial frames Agent Graphs as a response to “hardcoded” orchestration in frameworks such as LangGraph and OpenAI Agents, where routing logic and topology are embedded in code. Its pitch is that teams can see the whole workflow, swap a slow model from the dashboard, and reuse the same agent in multiple graphs. (launchdarkly.com 1) (launchdarkly.com 2) The result is a cleaner map of who handed work to whom, but not a full replay system. LaunchDarkly is offering a visual control layer for multi-agent workflows just as the wider tooling market is racing to capture the missing history behind agent failures. (launchdarkly.com) (arxiv.org)