In this June 7, 2026 post, Mark Dembo (Head of Solutions – Developer Platform & AI at Cloudflare) recounts a Sunday tinkering project: building a DIY AI agent from a small M5Stack Stick 3 device bought for ~€30, inspired by Steve Ruiz and motivated by the release of Opus 4.8. The motto: "There is no goal. Pure exploration and curiosity. And that feels just great."

First iteration (45 minutes). He throws the device's documentation at Claude Code, which produces roughly 200 lines of Python scripts. Since they have "zero blast radius", he allows himself to not worry about the code at all. The flow stays human-in-the-loop (not very "2026" for his taste): Opus generates, he copy-pastes into the UIFlow web interface, runs it, and reports the result back. His role shrinks to three functions — idea generator, executor and judge. A Cloudflare Workers + Workers AI backend quickly adds text-to-speech, push-to-talk (speech recognition), and a central small LLM: the object listens, answers, and tells bad jokes.

Second iteration. Aiming for a real agent, he points Opus at the Cloudflare Agents SDK, switches from REST to WebSocket transport, and enables Dynamic Worker execution — unlocking his favorite pattern, "Code Mode": the agent writes and executes code to accomplish its task. Given internet access, the agent computes 11! with a one-liner, finds the Champions League winner via fetch() on Wikipedia, and gives the weather for any city. Its limit: no access to private data.

Third iteration. He connects Todoist via an MCP OAuth flow (added in a few minutes by Claude) — and inherits 50 tools all at once, hence two problems: context bloat and a real blast radius (a bad call could delete a critical task). His countermeasure, inspired by Cloudflare's MCP Server Portal and Claude connectors: set each tool to Always allow / Ask for approval / Disable — disabled tools never enter the context, an LLM classifier only accepts explicit grants, and default = deny.

What he did not do: no latency optimization, no evals, no optimistic calls — "I did not even look at the code once. And you know what? That's the nice part of it." The final sense of wonder is less about capability than about its accessibility: €30 and a few cents of inference for an object controlled in natural language.