Cloudflare Adds Agent Tracing, with Truncation Limits and Uneven Payload Defaults

Details

Cloudflare's agent tracing extends existing Workers traces with agent-specific spans: invoke_agent, chat (model call with token usage), execute_tool, and tool_approval. Subagent calls nest under their parent in a single waterfall view, letting developers see infrastructure and agent behavior together for the first time. Three identifiers tie spans to the dashboard: agent name (the logical implementation class), agent ID (the running instance), and conversation ID. Cloudflare warns against deriving agent names from request-level data, which would fragment the view unnecessarily. Session replay reassembles recorded conversations turn-by-turn without re-executing them. A meaningful limitation: the tool_approval span records lifecycle events within a single Worker invocation, not how long a human actually waited before approving. Human-in-the-loop latency — arguably the most actionable debugging metric for approval workflows — is absent from the span. Developers using Cloudflare Agents now get structured debugging for agent behavior that was previously invisible, but need to handle human-approval timing separately.