SOURCE

Introducing: Cloudflare Agents

Properties1
tagscloudflare, agents, observability, tracing, primary-source

2 min read

Introducing: Cloudflare Agents

Summary

Cloudflare Agents begins as an observability product for deployed agents. Agent-aware traces connect model calls, tool executions, approvals, subagent work, and token usage to underlying Workers operations such as D1, KV, Durable Objects, service bindings, and fetches. This matters because an agent can return a technically successful response while choosing the wrong tool, retrying wastefully, using stale context, or failing the user’s actual job.

The dashboard provides two complementary views. Session replay reconstructs recorded messages, reasoning, tool arguments and results, and handoffs; it is a replay of stored data, not a re-execution. Execution traces show a waterfall of agent and infrastructure spans, including latency, model selection, token usage, tool calls, and nested subagents. Payload recording can be disabled where prompts or tool data may contain secrets or personal information. Supported launch integrations include Think, Flue, and AI SDK; custom harnesses can emit compatible spans, and traces can be exported to OpenTelemetry-compatible destinations.

For the proof-business funnel, the immediate value is operational rather than customer-facing. Mystery-shopper and AI-visibility runs can retain a defensible record of what happened, what it cost, where it failed, and whether a person approved a consequential step. Structured traces can also feed evaluations and cost reporting, helping repeated manual work reveal what is ready to productize. The announcement does not itself prove that Cloudflare Agents is the best or cheapest stack for an initial prototype, and tracing is beta-priced: free at publication, with Workers Observability pricing scheduled for October 1, 2026.

Key Claims

  • Agent traces cover model calls, tools, approvals, supported subagents, token usage, and connected Workers operations.
  • Session replay shows recorded context and tool activity without re-running the agent.
  • Payload capture is configurable for sensitive-data boundaries.
  • Trace export uses OpenTelemetry-compatible destinations.
  • Tracing was beta/free at publication; future pricing and retention were explicitly published.

Connections

Contradictions

  • None. It replaces the roundup’s general “watch every run” language with concrete scope and maturity details.