Cloudflare has rolled out agent tracing, a new dashboard feature that collects sessions from deployed AI agents in a single location, according to an announcement published by InfoQ. The release extends the company's existing Workers tracing by adding agent-level spans that capture model invocations, tool execution, and approval workflows—elements that infrastructure logs alone can't reveal. The feature addresses a core problem: an agent can return a successful HTTP 200 status code while still failing because it selected the wrong tool, passed outdated context to a subagent, or wasted tokens in a retry loop.
Agent tracing instruments each turn of an agent's operation as a nested waterfall, starting with the agent invocation and branching into chat model calls, tool execution, and approval steps. When a parent agent hands off work to a subagent that queries a database and writes to key-value storage, the entire sequence appears as a single trace spanning both infrastructure and agent layers. Three identifiers link spans to the dashboard view: an agent name for the logical implementation, an agent ID for the specific instance, and a conversation ID. Cloudflare cautions against deriving the agent name from a request or user identifier, which would fragment distinct agents in the interface. Session replay accompanies the trace view, reassembling recorded conversations across turns with messages, reasoning, tool calls with arguments and results, and subagent activity—though Cloudflare states this replays recorded data rather than re-executing the agent.
The documentation flags several limitations that matter for teams treating the feature as more than a debugging aid. Traces aren't a complete or lossless record of a conversation, and payload data faces span size limits, meaning long messages, reasoning chains, tool arguments, and results may be truncated. Session replay doesn't display images, and approval spans capture lifecycle events within a Worker invocation but don't measure the time a person waits before responding across invocations—human-in-the-loop latency, which the report describes as arguably the most interesting number in an approval workflow, isn't what the span records. Payload recording defaults vary by harness: Think doesn't store message or tool payloads unless developers explicitly enable storeMessages and storeTools, while Flue stores messages, system instructions, tool definitions, arguments, and results by default unless teams set content to false. The pricing structure carries its own surprise: every span counts as one observability event, including spans from SDK internals and other Worker-level operations the dashboard doesn't surface, so a verbose harness costs more than the view suggests.
Cloudflare positions tracing as a step toward agents that improve themselves, feeding structured trace data into evaluations and an agent development lifecycle. That vision remains a stated direction rather than a current capability. What exists today is visibility into which model a turn invoked, how many tokens it consumed, which tool it selected, and where the time went—matching a pattern the report identifies across vendors this quarter, including Microsoft's Agent Framework harness with OpenTelemetry enabled by default and Azure API Management's AI Gateway tier exporting token metrics to Application Insights, Datadog, and Grafana. From October 1, 2026, tracing will fall under Workers Observability pricing: the free tier covers 200,000 events per day with three-day retention, while the paid tier includes 20 million per month with seven-day retention at $0.60 per additional million. The takeaway is clear: infrastructure telemetry alone can't explain what an agent did, and teams need a dedicated layer to see the model behavior that drives the API calls and database queries underneath. For organizations running agents in production, the decision to instrument now—while tracing remains free during beta—depends less on whether agent-level visibility matters and more on whether the payload recording defaults match their data governance policies before the first trace goes live.