From Chatty Assistants to Compiler-Style AI Coding Agents
AI coding agents are automated systems that turn natural language or specifications into running code, and a new wave of compiler-based AI tools is rejecting interactive feedback loops in favor of deterministic, self-contained execution that returns complete services instead of partial drafts for human review. For years, AI coding tools have centered on developer feedback loops: prompt, inspect, tweak, repeat. Their value story is speed and convenience, helping programmers write code faster but keeping them in the decision-making path. Now leading voices in programming language design argue that this model attacks symptoms rather than causes. They say correctness problems stem from architecture, state and concurrency, not from how quickly developers can edit AI-generated snippets. This shift sets up a clash between conversational assistants that expect constant feedback and compiler-like agents that aim to ship validated systems with minimal human intervention.
Andrew Kelley’s Rejection of Subscription AI Coding Models
Andrew Kelley, creator and BDFL of the Zig programming language, has become a high-profile skeptic of mainstream AI coding agents. Zig’s code of conduct bans AI-generated contributions because Kelley says they are “invariably garbage” and waste reviewer attention that could go to mentoring human contributors. He also objects to non-deterministic AI tooling: if a simple refactor might behave differently from run to run, developers must inspect everything, which he sees as the opposite of reliable automation. His criticism reaches beyond quality into economics and control. Kelley refuses to rely on remote, closed AI services that require ongoing subscriptions, arguing that moving from local tools to “closed-source programming on someone else’s computer through the network” that he must pay for monthly is an “insane proposition.” In his view, developer tools should be deterministic, locally controlled, and held to a standard of “uncompromising perfection,” not “surprisingly well.”
SkipLabs’ Skipper: A Closed-Loop, Compiler-Based AI Tool
Julien Verlaguet, creator of Facebook’s Hack language and founder of SkipLabs, argues that current AI coding agents solve the wrong problem by optimising the feedback cycle rather than the architecture. He calls building correct software “an architecture problem disguised as a coding problem” and says AI made that gap more urgent. SkipLabs’ new agent, Skipper, embodies a different theory of AI coding tools: a closed-loop system that behaves more like a compiler than a chat assistant. Developers provide a plain-language description or OpenAPI spec, and Skipper returns a running backend service with routes, data mappers, validators, TypeScript types, unit tests, and a Docker container. The review-and-refine loop runs internally up to a fixed number of attempts, without back-and-forth prompts. According to The New Stack, Skipper “sits between foundation models and shipped software,” treating large models as commodity APIs while the Skip-derived reactive runtime manages state, cache invalidation, and concurrency.
Why These Agents Don’t Want Your Feedback
The core bet behind compiler-based AI coding agents is that feedback-driven iteration cannot reliably repair architectural mistakes. Tools like Claude Code or Cursor rely on developer feedback loops: humans catch subtle state and concurrency bugs, then prompt the model to patch them. Verlaguet counters that the real breakage happens where AI models have the weakest reasoning power: long-lived state and concurrent updates. Skipper avoids asking the model to juggle mutable state directly; instead, the Skip runtime defines programs as declarative graphs from inputs to outputs, handling caching and concurrency behind the scenes. Kelley’s perspective points in a similar direction: if AI output is non-deterministic and permanently needs checking, it fails as an automation tool. Both stances push against the idea that better prompting will fix systemic issues, arguing that architecture and runtime design must do more of the heavy lifting than interactive feedback sessions.
Developer Workflow or Developer Replacement?
Beneath these design choices is a sharper question: should AI coding agents augment workflows or make developer involvement optional? Skipper’s closed-loop design assumes that, for many backend services, the human’s main job is to describe requirements, not to shepherd code through generation and review. Kelley, by contrast, wants tools that preserve human authority while behaving predictably, more like advanced compilers than probabilistic co-pilots. Both views resist the “vibe coding” trend, where developers offload entire tasks to opaque systems controlled by a few large providers. For teams, this tension will shape how they choose between feedback-heavy assistants and compiler-based AI tools. Some will keep interactive agents alongside traditional compilers for exploratory work. Others may adopt closed-loop agents for well-bounded services, relying on deterministic runtimes and strong programming language design to keep shipped systems reliable while reducing the cognitive cost of babysitting AI-generated code.
