Cloudflare Agents Gain Observability

Cloudflare is making it easier to deploy and manage AI agents on its platform with the introduction of Cloudflare Agents. The company is consolidating agent operations into a single experience, starting with enhanced observability tools.

For developers building AI agents, understanding what's happening under the hood is critical. Agents can appear to succeed at the HTTP level while failing internally. They might select the wrong tools, pass stale data, or get stuck in retry loops. Traditional telemetry often misses these nuanced agent-specific issues. Cloudflare's new agent tracing aims to close this gap.

Deeper Insights into Agent Behavior

The new agent tracing feature offers granular visibility into an agent's actions. It meticulously tracks every model call, tool execution, and token usage. This allows developers to pinpoint where time is spent, whether it's in model inference, tool interaction, or underlying infrastructure. Traces will detail which models were called, how many tokens were consumed, and the success or failure of external API calls made by tools.

This agent-level telemetry goes beyond existing Workers tracing, which focuses on infrastructure spans like fetch calls or database queries. Agent tracing adds specific spans for agent invocations, model calls, tool executions, and approval events. Metadata such as model and token usage is attached, providing a comprehensive view of the agent's decision-making process. The initial rollout supports OpenTelemetry-compatible agent harnesses like Think, Flue, and AI SDK, with more to follow.

All Agents, One Dashboard

Cloudflare's dashboard now features a dedicated 'Agents' view. This centralizes all deployed agents, showing their traces, runs, sessions, and token consumption. Developers can visualize and debug agent behavior by replaying sessions or inspecting traces. Session replays reconstruct the full conversation, including system instructions, user messages, model reasoning, tool calls and their results, and the final output. This allows for identifying issues like malformed tool arguments or understanding how context influences later decisions. Importantly, payload recording for messages and tools can be optionally disabled to protect sensitive data.

The Traces tab provides a waterfall view, connecting agent operations directly to the underlying Workers infrastructure. For example, a 'Travel_Planner' agent might delegate to an 'itinerary_builder' subagent, which then calls a model, queries a database (D1), and writes to key-value storage (KV). All these steps, including their durations and associated token usage, are visible in a single trace.

The Road to Self-Improving Agents

Cloudflare's vision for Agents extends beyond mere deployment and observation. By integrating this detailed telemetry data back into the agent development lifecycle, the company aims to facilitate the creation of autonomous, self-improving agents. This capability could significantly accelerate the iteration cycle for complex AI applications. For startups and enterprises alike, having a unified platform to deploy, monitor, and refine AI agents offers a substantial advantage in building sophisticated AI-powered products and services.