Multi-agent AI systems introduce observability challenges that traditional monitoring tools weren't designed to handle. When agents orchestrate other agents, failures can propagate silently through the chain without triggering any errors. Proper observability requires trace continuity across agent handoffs, per-agent span attribution for latency and token usage, differentiated failure modes, and cost attribution at the task level. Sentry's distributed tracing primitives can be applied to multi-agent pipelines, with auto-instrumentation available for major AI frameworks including OpenAI, Anthropic, LangChain, LangGraph, and others in Python and Node.js. The post recommends starting by instrumenting the orchestrator first, capturing each agent call as a child span, and extending with custom spans for unsupported frameworks.
