
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Cloudflare thinks AI agents need new primitives, not just wrappers — Sunil Pai argues that Durable Objects plus Dynamic Workers give Cloudflare a real architectural edge for long-running agents, with stateful serverless actors and safe execution of user- or LLM-generated code at zero startup time.
Dynamic Workers are Cloudflare’s bet on safe code execution as a core agent capability — Pai says you can expose only specific APIs, block outgoing traffic by default, and let models submit JavaScript that searches or executes across Cloudflare’s 2,600 API endpoints without exploding into thousands of tools.
His favorite example is an MCP server that collapses massive API surfaces into two tool calls — Instead of one tool per endpoint, Cloudflare’s MCP setup uses just “search” and “execute,” letting an LLM write code like “find all my workers tied to DNS zones starting with Z and apply DDoS protection” in a single typed operation.
Pai sees today’s agent platforms as pre-React: everyone is building harnesses, but nobody has built the breakthrough abstraction yet — He compares the current moment to 2013 React skepticism and says the eventual standard must be reproducible across languages, companies, and infrastructure, with “skills” or even Markdown-like interfaces as plausible contenders.
The ‘slop fork’ Vercel drama became a bigger story about open source culture than corporate rivalry — Pai says he casually ported Vercel Labs’ Just Bash to Cloudflare with Opus generating 5,000 lines of code, expected to polish it and send a PR, then woke up to Vercel’s CTO criticizing it publicly and half the internet defending him.
His closing message is blunt: stop making incremental AI products and build ‘science fiction’ instead — Pai urges builders to use the new leverage from LLMs and infrastructure to try weird, original ideas for real people, not “another agent framework” or a safer path to an enterprise deal.
The conversation opens with Anthropic’s newly launched cloud-managed agents, which Pai says is “dope” because it treats long-running agents as a real platform problem. But his immediate reaction was competitive: he looked at it and thought Cloudflare could build something better with Workers and Durable Objects.
Pai says the industry is still “discovering the architecture of the new architecture of building software,” and he centers two Cloudflare primitives. First is Durable Objects, which he calls the infrastructure-layer implementation of the actor model: millions of stateful serverless entities without spinning up full VMs. Second is Dynamic Workers, which safely run eval-style code with zero startup time while tightly controlling available APIs and outbound traffic.
His most concrete example is Cloudflare’s MCP server, built by colleague Matt Carry, sitting on top of an API with 2,600 endpoints. Instead of turning every endpoint into a tool, they expose only “search” and “execute,” and let the LLM send JavaScript into an isolate to inspect the OpenAPI spec or perform actions. Pai’s pitch is that models are actually great at this kind of code-running workflow, so you can compress huge capability surfaces into a couple of typed operations.
Asked whether this kind of harness architecture becomes standardized, Pai says the field is still too early and “no one has built the React yet.” He reaches back to 2013, when people literally walked out of React talks at Facebook, to make the point that the winning abstraction may look weird at first. His current hunch is that “skills” might be the most portable translation layer because they scale, are English-addressable, and can still bundle code.
Pai then recounts the controversy around porting Vercel Labs’ Just Bash project to Cloudflare. While in Spain at JSConf, he says he pointed Opus at the repo, went to lunch, came back to 5,000 generated lines, and planned to clean it up and submit a proper PR after recovering from travel. Instead, he woke up from a nap to DMs from Cloudflare management and a public post from Vercel’s CTO, which he says made him briefly feel like he was watching his career fall apart.
Pai insists forking is not theft in his culture — it’s prestige, respect, and how communities like React actually grew. He says people should absolutely fork the Agents SDK if they think they can do better, because “everyone wins” when software gets remixed and improved. That leads to a broader critique: today’s open-source repos have become adversarial, with fake security reports, noisy contributions, and maintainers increasingly afraid of popularity.
The final note is classic Sunil Pai — energetic, idealistic, and a little exasperated. His ask to builders is simple: be original, be courageous, and use this weird new era of LLMs to build “sci-fi stuff,” not another incremental framework chasing enterprise buyers. In his words, infrastructure and models almost fail you at the edge right now — which is exactly why that’s where people should experiment.
Share
Keep Reading
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.

Playbook
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.

Playbook
OpenAI shipped /goal for the Codex CLI. It turns a prompt into a persisted, self-continuing contract.