Cloudflare devs @ AI Engineers Europe (Sunil Pai, Matt Carey & Thomas Ankcorn)
Cloudflare engineers Sunil Pai, Matt Carey and Thomas Ankcorn make the case for radical simplicity in AI tools: let models write code, keep the interface small, and leave room for rapidly improving models to do more of the work.
Recorded at AI Engineer Europe, this conversation brings together three members of Cloudflare's developer team: Sunil Pai, Matt Carey and Thomas Ankcorn.
They talk about minimal coding agents, persistent home-lab sandboxes, Code Mode, the connection between MCP and observability, and why the most ambitious experiments can begin as deliberately dumb prototypes.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
What stood out at AI Engineer Europe?
The density of builders made the conference unusually useful: people recognized from open-source projects and online work were all in the same place. Beyond the talks, Sunil argued that high-quality industry gossip is a leading indicator that the right, deeply connected people have shown up.
Which coding-agent ideas were you excited about?
Thomas liked Pi's speed, tiny surface area and extensibility, alongside a project that exposes a persistent home Raspberry Pi to an agent through MCP. The appeal is continuity: the sandbox survives when the user changes machines. Matt highlighted the unusually good terminal interface and how little machinery the agent needs.
What is Code Mode?
Instead of exposing a growing collection of static tool calls, give the model one environment where it can write code against an API or external service. Modern foundation models are good enough at code that this can be more capable and context-efficient than describing every operation as a separate tool.
Why might code become an interface for far more users?
Programmers already automate a repetitive request — such as captioning hundreds of photos — by writing a script. Agents can extend that leverage to people who would never open an editor. The new infrastructure question is where arbitrary generated code can run safely, because the industry spent decades teaching systems not to evaluate it.
How are coding agents changing work across a large codebase?
Observability work at Cloudflare touches runtimes, control planes and data repositories. Agents make it much easier to prepare changes across those boundaries, so the harder task becomes earning maintainers' trust and building the human relationships that get good changes reviewed and merged.
How do Code Mode, MCP and observability fit together?
They can close an operational loop: an agent observes a system, accesses capabilities through MCP, and writes code to adjust or repair it. At Cloudflare's scale, that means building infrastructure able to understand and respond to its own planetary volume without a person manually tracing every configuration mistake.
Why should founders build deliberately simple prototypes?
Remove complexity, build the dumb version and put it in front of people. Models improve every few months, so a heavily engineered workaround can become obsolete before it pays back. Simpler, interpretable systems benefit more from the next model generation — and leave founders free to attempt stranger, more original products than another incremental SaaS application.