Julien Verlaguet, founder of SkipLabs, has spent the last year grappling with this question, and the answers he’s uncovered are far from satisfactory. "Every time I look closer at people who claim that they are bringing guardrails to AI, I see more of the same," Verlaguet told The New Stack. "I see more prompting – and I don’t see anybody who is trying to build real guardrails and real tooling from scratch." His explanation for this deficit is blunt: "It’s a lot of work, and so it’s much easier to make those big claims that you’re going to do these things when, in the end, you don’t."
Verlaguet, however, is committed to doing that work. SkipLabs is developing Skipper, a specialized coding agent designed to generate and maintain backend services. This is not a direct code generator in the vein of tools like Microsoft’s Copilot. Instead, Skipper focuses on the underlying structural layer of AI-assisted development, aiming to make AI-generated output readable, maintainable, and deployable at speed.
"The first thing to notice is that Skipper is not a model," Verlaguet clarified. "So, we’re not doing any AI here. We treat models as a commodity. We use different models, Anthropic for the most part, but not only, but to us, the model is just an API that we call with a context, and it comes back with the result."
Verlaguet’s core thesis for this approach begins with a critical latency problem. "In the next few years, it’s not going to be okay to wait for CircleCI to take half an hour to an hour to validate a diff," he asserted. "AI is getting faster and faster, and so you will need the tooling to guardrail the AI, and that tooling will need to be incremental."
The Incremental Advantage: A Legacy of Precision
Verlaguet’s career has been built on the principle of incremental development. At Facebook, he was instrumental in the creation of Hack, a gradually-typed dialect of PHP. This innovation introduced a type system to a dynamic language years before TypeScript popularized a similar approach. Hugo Venturini, a software engineer at SkipLabs, elaborated on this in a blog post, explaining that Hack was developed to address the fundamental type-unsafety of PHP, which powered the entire Facebook codebase. At Facebook’s scale, this was not merely an aesthetic issue but a significant engineering liability.
Venturini wrote, "So, Julien built Hack: a gradually typed, rigorously annotated replacement that was strictly less pleasant to write than PHP. It was more verbose. It demanded precision. It required you to say exactly what you meant. And then they got every engineer at Facebook to switch." This transition underscored Verlaguet’s ability to drive adoption of more precise and robust development practices, even when they demanded a steeper learning curve.
Following Hack, Verlaguet developed Skip, a full reactive programming environment. The core principle of Skip was that when inputs change, there should be no need to recompute everything from scratch. Observing a lack of similar frameworks outside of Meta, Verlaguet and his team founded SkipLabs to bring this reactivity paradigm to a broader developer community. Lucas Hosseini, another software engineer at SkipLabs, described reactive programming as "a declarative way of expressing computations: instead of manually handling state transitions, you simply describe state as a function of multiple inputs."
Verlaguet departed Meta in 2020 to focus on his startup, and Skipper represents the culmination of his work in reactive technology.
Two Parts, One Stack: The Architecture of Skipper
Skipper is composed of two primary components. The first is a development environment built upon a novel, sound, and incremental implementation of TypeScript. Verlaguet’s deliberate choice of TypeScript was strategic, as he noted, "AI is very good at TypeScript and Python. I think starting with one of these two is probably where you’re going to get the best results."
The technical soundness of the type system is paramount. A type system with inherent "holes" cannot support the reachability analysis – the call-graph mapping that determines which code changes affect which program state – which is central to Skipper’s functionality. "If the type system wasn’t sound, then you don’t know what the types are, you don’t know what you’re calling, you cannot build a call graph," Verlaguet explained. A reactive runtime then sits atop this foundation, updating live state as code changes occur, thus eliminating the need for program restarts.
The second component of Skipper is the harness: the agentic orchestration layer. While structurally classic in its approach (plan, generate tests, generate code), it is built on the same incremental framework. This allows for generation and remediation to occur in parallel on independent code segments, rather than sequentially across an entire codebase. The combined effect is that Skipper can ingest code changes at high speed, re-run only the tests impacted by those changes, and update a live service without any downtime.
Beating Claude Code at Its Own Game: Performance Benchmarks
In internal benchmarks, Verlaguet claims Skipper successfully passes over 90% of tests within their corpus of backend service prompts. This significantly outperforms Claude Code, which achieves approximately 20% on the same dataset. However, Verlaguet is careful to define the scope of this comparison. "Claude Code does a lot more," he acknowledged. "We are better at what we do." What SkipLabs focuses on is generating general-purpose backend services, not relying on a catalog of predefined templates. "All the code is generated from scratch," he emphasized.
Verlaguet posits that AI will increasingly drive software development towards service architectures. This shift, he believes, is not merely a trend but a necessity driven by the fact that stateful services are essential for making iterative AI development tractable. He uses the analogy of a compiler: "In five years from now, I’m pretty sure you will want this compiler to be built by an AI. And what does that mean? It means that every time the AI wants to iterate on that compiler, it has to wait half an hour. It doesn’t make any sense." By transforming the compiler into a stateful service, and sending requests to it, AI can iterate without such prohibitive overhead.
While Skipper is still in its nascent stages, SkipLabs is reportedly poised for an announcement in the coming month. The current user base is described as "us and our friends," indicating a focus on internal testing and refinement before a broader release.
Verlaguet characterizes SkipLabs as a specialized coding agent shop, a category he anticipates will become increasingly competitive. "I think we’re going to see different agents with different tools, tool chains, that are better at doing certain things," he predicted. The crucial differentiator, he argues, will be the presence of genuine guardrails, distinguishing substantial solutions from mere noise. "Here’s a company who’s actually built the guardrails you were looking for, for your AI-generated code," Verlaguet stated.
Industry Recognition and Emerging Issues
The technical ambition of SkipLabs has not gone unnoticed by industry observers. Brad Shimmin, an analyst at The Futurum Group, described SkipLabs’ technical creation as "Very fascinating and a reflection of how software is changing in our current non-deterministic world. Instead of using traditional but sometimes network-heavy declarative frameworks for real-time response, this framework and back-end service basically use a declarative mechanism to just reason over what a code block is supposed to do, tracking any dependencies in a computational graph."
David Mytton, CEO of AI security platform provider Arcjet, echoed this sentiment, stating that Skipper addresses a critical and emerging issue. "I’ve had multiple conversations with technical leaders recently and the industry is still coming to terms with the death of code review," Mytton told The New Stack. "We’re in a new world where the old security model breaks down because it assumes a human wrote the code. Review can’t keep up when agents handle implementation – and they’ll soon manage the full cycle from planning through to deployment. Security must be baked into the whole process, from the development environment to runtime."
The End of the Readability Era: A Machine-Native Future
In his blog post, Venturini argued that "An agent benefits from verbosity. Long, precise, unambiguous tool outputs are easier to parse than short, clever, human-optimized ones." He further contended that loud, specific failures are more valuable than errors that are masked for the sake of developer experience. Strong type systems, in this view, serve not as mere guardrails but as the most information-dense descriptions of what code is intended to do.
This perspective points towards a bifurcated tooling landscape. Venturini suggests that languages and environments designed for human authors will persist, as "humans aren’t going anywhere." However, alongside these, a new generation of tools is emerging with the AI agent as the primary consumer: tools that are strict, formally specified, and intolerant of ambiguity.
Skip Labs positions Skipper squarely within this second category. Its TypeScript-compatible type checker, SKJS, is sound in ways that standard TypeScript is not. This trade-off, while potentially making the system less user-friendly for humans, enhances its utility for AI agents. The reactive runtime enforces explicit dependency contracts that might seem excessive to a human developer but are precisely what an agent requires to reason about cause and effect within a codebase.
The underlying argument is that if AI agents are responsible for generating the majority of code, and the tools they utilize were originally designed for human readability, a significant capability ceiling is being imposed. It’s akin to asking an agent to operate within a medium not designed for its inherent strengths. "We made the tools more readable to get more developers," Venturini wrote. "We’ll make them less readable – more precise, more formal, more machine-native – to get better agents. The readability era of programming languages was long and productive and is now coming to an end."
A Timeline of Innovation
Julien Verlaguet’s journey leading to the development of Skipper and the potential product release is marked by significant milestones in software engineering:
- 2011-2015: While at Facebook, Verlaguet played a pivotal role in the creation of Hack, a gradually-typed dialect of PHP. This initiative introduced type safety to a large, dynamic codebase, laying the groundwork for his later work on precise development tools.
- 2015-2019: Verlaguet developed Skip, a reactive programming environment designed for efficient state management and incremental updates. This project explored principles of functional reactivity and its application in complex software systems.
- 2020: Verlaguet departed Meta to found SkipLabs, with the aim of commercializing and expanding the reach of reactive programming technologies.
- 2021-Present: SkipLabs focused on building Skipper, a specialized coding agent and development environment. This period involved intensive research and development into incremental computation, sound type systems, and agentic orchestration for backend service generation.
- 2024 (Expected): SkipLabs is anticipated to make a significant product announcement regarding Skipper, signaling its readiness for wider adoption.
The trajectory of Verlaguet’s work highlights a consistent drive towards enhancing the robustness, maintainability, and efficiency of software development, particularly in anticipation of the profound impact of AI on the industry. The development of Skipper represents a direct response to the perceived shortcomings of current AI development tools and a commitment to building the foundational "guardrails" that the industry widely acknowledges as necessary.