When AI Agents Orchestrate Agents: The Observability Gap
When AI agents spawn other agents, failures often stay invisible. Sentry's trace-based observability approach reveals exactly which agent introduced the problem.
AI systems no longer mean a single model call — orchestrator agents now spawn sub-agents for planning, retrieval, validation, and execution. In these architectures, failures rarely surface as clean exceptions; instead they appear as silent context corruption that propagates down the chain. The orchestrator may report success while the user sees a broken result, with nothing obviously wrong in traditional logs.
When five agents call each other conditionally with shared context, the problem shifts from debugging code execution to debugging emergent behavior across a distributed decision graph. Just as microservices made 'it's slow somewhere' meaningless without tracing, multi-agent systems make 'the AI did something wrong' unactionable without proper instrumentation.
Effective observability requires trace continuity across agent handoffs, per-agent span attribution for latency/tokens/model version, differentiation between failure modes (timeouts, bad tool outputs, context overflow, hallucination), and task-level cost/token attribution. Sentry extends its existing distributed tracing primitives with auto-instrumentation for major frameworks like OpenAI, Anthropic, and LangGraph, while unsupported frameworks need only about ten lines of manual span code.
Sentry's Seer tool helps narrow down the likely failure source across a five-agent trace, cutting triage time. The practical starting point for engineers: instrument the orchestrator as a top-level transaction first, with each agent call as a child span.