INFRA Signal 583
Introducing: Cloudflare Agents
↗
Cloudflare has launched a unified Agents experience in its dashboard, beginning with agent-aware distributed tracing that adds spans for model calls, tool executions, approval events, and subagent activity on top of existing Workers infrastructure spans.
For engineers running agentic workloads on Workers, the new view fills a real observability gap: prior traces covered fetch, KV, and D1 operations but stopped at the agent boundary, leaving you unable to tell whether a slow turn was the model, a tool, or the network. First-day support is limited to three OpenTelemetry-compatible harnesses, so adoption cost depends on whether your stack already uses one of them. Replay is recorded rather than re-executed, and payload capture is opt-in for sensitive data, so teams will need to make explicit choices about what gets stored.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Cloudflare's new agent tracing layers agent-operation spans on top of existing Workers infrastructure spans, with model name and token usage attached as trace metadata.
Initial harness support is limited to Think, Flue, and the AI SDK, with traces viewable in the Cloudflare dashboard or exported to an OpenTelemetry-compatible destination.
The Agents dashboard view includes a session replay of recorded messages and tool payloads, gated by storeMessages and storeTools flags that can be disabled when payloads contain sensitive data.
THE CLUSTER