Cloudflare Launches Agent Tracing to Debug AI Workflows

Cloudflare has launched agent tracing, the inaugural feature of its new Cloudflare Agents dashboard designed to centralize deployed agent sessions. While traditional infrastructure monitoring captures standard database queries and API requests, it often misses the internal logic failures of AI agents, such as looping retries or incorrect tool selection. The new system addresses this by adding agent-level spans for model calls, tool executions, subagent invocations, and approvals directly on top of existing Workers tracing.

The system groups each interaction turn into a single trace, nesting subagent activities under their parent operations. Developers using the Think framework or Flue version 2 and later will benefit from automatic instrumentation. Those making direct AI SDK calls can use the wrapAISDK wrapper, which supports versions 6 and 7. For custom setups, developers must use the Workers custom spans API, which aligns with OpenTelemetry Generative AI semantic conventions and exports to any OTLP endpoint.

Practitioners must navigate inconsistent payload storage defaults across different integrations. While Flue automatically records system instructions, messages, and tool definitions unless disabled with a content: false setting, Think and wrapAISDK require explicit activation via storeMessages and storeTools. Furthermore, Cloudflare notes that traces are not lossless; long messages, reasoning steps, and tool arguments are subject to truncation due to span size limits, and session replays do not display images.

The tracing feature is free during its beta period, after which it will transition to Workers Observability pricing on October 1, 2026. Under this model, every span counts as an individual observability event. The Workers Free tier will cover 200,000 events daily with a three-day retention period. The Workers Paid tier will include 20 million monthly events with seven-day retention, charging $0.60 per additional million events thereafter.