I’ve been working on AgentPulse, an open-source, self-hosted tool for detecting and investigating behavioral drift in multi-agent systems.
LangChain and LangGraph felt like a natural fit because AgentPulse works with the same structure developers already think in: agents, nodes, tool calls, handoffs, and execution paths.
The problem I’m exploring is when a workflow still completes, but its behavior has quietly changed. An agent may start using different tools, taking longer, costing more, or affecting downstream nodes without producing an obvious error.
AgentPulse compares runs and versions, flags drift across agents, handoffs, and routes, and brings together the traces and recent prompt, model, or tool changes that may explain it.
It’s still early, and I’d genuinely appreciate feedback from people building with LangChain or LangGraph, especially on whether this reflects a real problem in your current workflow.
Website: AgentPulse: drift investigation for multi-agent systems