Cloudflare Agents Enhance AI Observability and Management
Cloudflare launches Cloudflare Agents, a new platform for deploying and managing hosted AI agents, starting with advanced observability features.
Cloudflare has introduced Cloudflare Agents, a new platform designed to streamline the deployment and management of hosted AI agents. This initiative marks a significant step in Cloudflare's commitment to providing a robust developer platform, extending its existing infrastructure to accommodate the unique needs of AI agent development.
The core of the new offering is enhanced observability, starting with agent tracing. This feature provides developers with granular insights into agent behavior, meticulously tracking every model call, tool execution, and token usage. This detailed telemetry aims to empower developers to better understand, debug, and optimize their AI agents by offering agent-specific insights that complement Cloudflare Workers' existing tracing capabilities.
Agent tracing is designed to answer critical questions about an agent's performance and resource consumption. Developers can now pinpoint where time is spent—whether in model inference, tool execution, or underlying infrastructure. The system tracks specific model calls, token counts, tool selections, and the outcomes of external API interactions initiated by agents. This level of detail is crucial for diagnosing issues, such as an agent choosing an incorrect tool or getting stuck in a retry loop, which might appear as a successful API call to traditional monitoring systems.
Cloudflare Agents integrates agent-level telemetry with its existing Workers tracing. While Workers tracing already covers the infrastructure layer (e.g., fetch calls, KV reads, D1 queries), agent tracing adds a layer of visibility into the agent's operational logic. This includes spans for agent invocations, model interactions, tool executions, and approval events, all correlated with the underlying infrastructure data. Contextual metadata, such as model names and token usage, is attached to these spans.
Support for popular OpenTelemetry-compatible agent harnesses, including Think, Flue, and AI SDK, is available from launch. Agents built with these frameworks can automatically send their agent traces to Cloudflare, where they can be visualized directly in the dashboard or exported to other compatible OpenTelemetry destinations. This interoperability ensures that developers can leverage their existing tooling and workflows.
The Cloudflare dashboard now features a dedicated 'Agents' view, consolidating all observed agents and their associated traces. This central hub allows users to review agent runs, sessions, instances, and token usage. Within an agent's view, developers can replay sessions to examine the context across all conversational turns, including system instructions, user messages, model reasoning, tool calls, and final responses. This replay functionality is invaluable for identifying errors in logic or data handling without re-executing the agent.
Complementing the session replay is the 'Traces' tab, which presents a waterfall view of execution. This visual representation helps developers understand time distribution and correlate agent operations with Cloudflare Workers infrastructure. For instance, a trace can show a parent agent delegating tasks to a subagent, which then interacts with models, databases (like D1), and key-value stores (like KV), all within a single, correlated view. This end-to-end visibility is key to optimizing complex agent workflows.
Cloudflare Agents aims to be a comprehensive solution for the entire agent development lifecycle. By providing deep observability and a centralized management experience, Cloudflare is enabling developers to build, deploy, and continuously improve sophisticated AI agents, fostering the development of more autonomous and self-improving AI systems.
Cloudflare is further enhancing its AI agent capabilities by integrating local tracing directly into its Wrangler and Vite development tools. This new feature automatically captures OpenTelemetry traces for local Worker invocations, allowing coding agents to query these traces via the Local Explorer API for debugging. This significantly speeds up issue resolution by providing detailed insights into operations like KV reads or D1 inserts without requiring manual instrumentation or code changes.