The Dashboard Is the Wrong Interface
Technical product interfaces have barely evolved since the early web. Workflows still demand clicking through pages, correlating log trails across tabs, and discovering buried settings. Agent Lee is Cloudflare’s attempt to replace that with something fundamentally different: describe what you want to accomplish, in plain language, and let the platform handle the rest.
Agent Lee isn’t a chatbot bolted onto the dashboard; it’s an in-dashboard AI assistant with awareness of your entire Cloudflare account, including your Workers, zones, DNS configuration, and error rates. It can answer questions about this data, debug issues, apply configuration changes, and even deploy new resources on request.
Consider the classic 2 a.m. incident: a Worker returns 503s, and you have to hunt through R2, routes, and rate limits to find the culprit. Agent Lee understands these resources and can both diagnose the problem and fix it immediately, directly from a conversational prompt. It compiles real-time metric visualizations in the chat window, pulling from your actual traffic data rather than pushing you to a separate analytics view.
What’s Different Under the Hood
The agent’s capability also relies on two unusual architecture decisions. First, Cloudflare didn’t hand raw MCP tool definitions to the model. Instead, it uses its established Codemode approach to convert tools into a TypeScript API and has the model generate code against that interface. Because LLMs are more accurate with common TypeScript than with bespoke tool-call schemas, this improves reliability. For multi-step actions, code generation also allows the model to chain several calls into a single script, eliminating network round-trips and returning just the final outcome.
That generated code is then shipped to a Cloudflare MCP server for execution—but it goes through the agent’s permission system first. The system uses a Durable Object as a credentialed proxy that inspects the generated code and classifies it as read or write. Read operations are proxied silently. Write operations are blocked until you explicitly approve in the agent’s UI, and credentials which are crucial to the write are held server-side within the Durable Object, so they never appear in the static code being executed. That’s more than a sandbox; the permission gate is architectural.
Agent Lee connects to Cloudflare’s own MCP server, which exposes two primary tools: one for querying API endpoints and another for executing write operations. The write path requires a manual approval step from the user interface before the work goes out. That confirmation prompt is an enforced security layer, not a confirm-dialog courtesy.
Talking in Tables, Charts, and Diagrams
The UI isn’t stuck with plain text, either. As a conversation progresses, an adaptive grid renders next to the chat, allowing you to request specific dashboard widgets on the fly. Agent Lee responds by generating interactive tables, line charts, architecture maps, and other structured visualizations that show actual account data in real time. Instead of a wall of numbers from a top 5 errors query, your results arrive as a clickable, sortable chart.
Everything you ask for becomes a persistent, editable component in the grid, enabling you to tie data outputs together and essentially turn her conversation history into a personalized operations dashboard at runtime.
Built From Battle-Tested Azure
Agent Lee’s less-publicized but equally important story is that it’s constructed entirely from the same Cloudflare components any customer can access: the Agents SDK, Workers AI, Durable Objects, and the public MCP layer. Being deployed on its own primitives isn't just environmentalist purity; it means platform bugs are surfaced in real production scenarios, at substantial scale. Agent Lee currently processes roughly 250,000 tool calls daily for about 18,000 active users. Each one of those is a test case for Cloudflare’s underlying agent platform and permission architecture.
Production Observations
Because Agent Lee performs write operations against real estates, monitoring is heavily production-focused. Cloudflare reports it tracks a combination of signal types: conversation success rates and accuracy via evals, embedded thumbs-up/thumbs-down feedback, tool-call execution rates, hallucination scorers, and per-product performance metrics. This data feeds both the model and the permission framework, a loop that aims to tighten agent reliability over time without centralizing more unapproved power in the model.
Alpha Status and Next Horizons
Agent Lee is rolling out slowly. It is live now within the dashboard for Free plan users—click “Ask AI” in the upper right corner of the Cloudflare dashboard to try it. Cloudflare explicitly acknowledges that the beta is for iterative testing: you might hit edge cases or unresolved limitations, and user feedback shapes iteratively better features.
At a strategic level, the long-term road is to position her as the Cloudflare platform’s universal access surface—inside the dashboard, in your CLI, and on your phone. Current work is focused on ensuring Agent Lee moves from a system that reacts to prompts to one that watches specified metrics and surfaces conditions it deems relevant before you ask. But that proactive, context-aware future is dependent on the same foundational context Agent Lee has already gathered about your account configuration and action history.
