Cloudflare Agents Unveiled

Cloudflare has unveiled Cloudflare Agents, which unifies the deployment and operation of AI agents. It is a product that lets developers use the model access, durable runtime, orchestration, sandboxed execution, and persistent storage that Cloudflare's developer platform has built over the past 9 years to build agents.

Cloudflare Agents brings together all deployed agent sessions into a single environment, allowing teams to check agent performance and behavior throughout large-scale operations. As its first feature, it offers agent tracing, which captures agent behavior in fine detail.

Agent-aware traces measure and show the following information:

• Time taken and cost per model call
• Tool execution process and the results of tool selection
• Number of tokens used and which models were called
• Whether execution paused due to waiting for approval
• Whether external API calls succeeded, failed, or timed out
• Sub-agents' work and its effect on the final response

Existing application telemetry can track things like API requests or database queries, but it cannot surface problems that arise from an agent's actual behavior, such as choosing the wrong tool or passing along stale context. An agent can return an HTTP 200 response while still producing an incorrect result, or waste tokens in a retry loop.

Agent tracing supports agent harnesses that are compatible with OpenTelemetry, including Think, Flue, and AI SDK, with plans to expand support going forward. Cloudflare says it plans to connect observability data to the agent development lifecycle, building an environment for autonomous agent operations that enables deployment, observation, and continuous improvement.