Back to feed
News Story
APriority76
InfoQ AI/ML/Data Eng
1 sources

Cloudflare Launches Agent Tracing with Truncation Limits and Uneven Payload Defaults

Cloudflare has launched agent tracing, adding spans for agent invocations, model calls, tool runs, and approvals to existing Workers traces. Sessions replay turn by turn, though the docs warn traces are not lossless and payloads may be truncated. Payload recording defaults differ by framework, and from October 1, 2026 every span counts as a billable event.

SynthePulse Insight · AI deep readingMembers

Cloudflare Introduces Agent Tracing: Debugging Tool or Privacy Trap?

Version 1 · 1 source

Cloudflare has released Agent Tracing, providing observability for AI agents on Workers, but inconsistent default privacy settings, data truncation limits, and per-span pricing require developers to weigh carefully.

  • Cloudflare introduces Agent Tracing, adding spans for agent calls, model calls, tool executions, and approvals on Workers, with pricing under Workers Observability starting October 1, 2026.
  • The feature addresses the issue of agents returning HTTP 200 but actually failing, helping developers identify wrong tool choices, context passing errors, or retry loops by tracing agent behavior rather than just infrastructure.
  • Default payload recording policies are inconsistent: Think and wrapAISDK() do not store message and tool payloads by default, while Flue stores them by default, requiring explicit content: false to disable, posing potential privacy risks.
Open section navigationPositioning and Value of Agent Tracing

Positioning and Value of Agent Tracing

Cloudflare has launched Agent Tracing as the first component of Cloudflare Agents, aiming to centralize deployed agent sessions in a single dashboard. The feature adds agent-level spans to existing Workers tracing, covering agent calls, model calls, tool executions, and approvals, attaching model and token usage as metadata. Each turn generates a trace, with subagent work nested under the calling operation, forming a cross-layer waterfall view.

InfoQ's report points out that the feature solves the problem where agents may return HTTP 200 but still fail, such as selecting the wrong tool, passing stale context to subagents, or wasting tokens in retry loops. Traditional application telemetry only shows API requests or database queries, not the agent behavior causing the issue. Developer Mykyta Pavlenko responded on X, saying that seeing model calls directly above wrong tool parameters is exactly the debugging view he wanted.

However, there is a key limitation with approval spans: documentation states these spans represent lifecycle events within a Worker invocation and do not measure human wait time across calls. Therefore, the latency of human-in-the-loop—the most interesting metric in approval workflows—is not recorded.

Free for now

Read the full analysis

3 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This report is based on a single source from InfoQ, an industry media report, and has not been officially confirmed by Cloudflare. All facts are derived from that article, which includes a retelling of Cloudflare's announcement and developer comments.

Sources for this version

  1. Cloudflare Adds Agent Tracing, with Truncation Limits and Uneven Payload DefaultsInfoQ AI/ML/Data Eng

Primary report

InfoQ AI/ML/Data Eng

Primary source