← Field Notes
AUG 29 · Clipped · via The Changelog ApprovalObservabilityPermissions

A Cloudflare engineer gives agents code to run instead of a tool list

If the tool surface is generative, our permission and approval UI patterns assume the wrong thing. This is a foundational constraint worth surfacing before teams design those screens.

Machine summary of the source

A Cloudflare engineer working on the Agents SDK and MCP infrastructure describes a concrete alternative to the dominant MCP pattern of enumerating discrete tools: instead of registering a fixed set of named tools, you expose a code-execution environment and let the model write the tool itself at runtime. The distinction matters architecturally — tool enumeration forces you to anticipate every operation in advance and express it as a schema, while code-as-tool lets the agent compose novel operations from primitives without the server author having to predict them.

The interview surfaces real implementation detail: how the Cloudflare MCP server is structured, what the SDK makes easy versus hard, and where the seams are between the agent runtime and the tool surface. This is primary-source, practitioner-level signal — not a whitepaper, not a demo — which makes it useful for teams reasoning about what MCP server design decisions lock in downstream.

For UX, the downstream consequence is significant. If the tool surface is dynamic and generated rather than enumerated, the standard design assumption — that you can display, permission, and audit a bounded set of named operations — starts to break down. Interfaces built around 'here are the tools this agent can use' need a different model when the agent can synthesize tools on the fly.

The summary above is generated; the note at the top is the editorial judgment. Primary source ↗