Developer Builds Working Scheme Compiler in 4 Days Using Claude AI
Matthew Phillips' rapid construction of Puppy Scheme compiler demonstrates AI's accelerating capacity to tackle traditionally complex software engineering tasks.
Developer Matthew Phillips built a functional Scheme-to-WebAssembly compiler in just four days with AI assistance, a project that would typically require months of specialized compiler engineering work. The most striking demonstration of AI’s capabilities came when Phillips instructed Claude to ‘grind on performance’ before going to bed and woke up to find compilation time reduced from 3.5 minutes to 11 seconds—a 19x improvement achieved autonomously overnight.
The compiler supports 73% of the R5RS and R7RS Scheme standards, WASI 2, the WebAssembly Component Model, and uses WebAssembly garbage collection. Phillips deployed a working website running on the compiled WASM code in Cloudflare Workers, demonstrating production viability.
AI Agents Build Production Compilers
Phillips’ project parallels recent advances in autonomous AI software development. In February 2026, Anthropic researcher Nicholas Carlini deployed sixteen Claude Opus 4.6 agents to build a Rust-based C compiler from scratch that ultimately compiled the Linux 6.9 kernel across x86, ARM, and RISC-V architectures, according to InfoQ.
The autonomous C compiler project ran approximately 2,000 sessions and cost about $20,000 in API fees. It achieved a 99% pass rate on the GCC torture test suite and can compile PostgreSQL, SQLite, Redis, FFmpeg and Doom, as reported by Slashdot.
| Project | Developer | Time | Scope |
|---|---|---|---|
| Puppy Scheme | Matthew Phillips + Claude | 4 days | Scheme → WASM, 73% R5RS/R7RS |
| C Compiler | 16 Claude agents | 2 weeks | C → x86/ARM/RISC-V, Linux kernel |
| Darklang Compiler | Paul Biggar + Claude Code | 2 weeks | Full optimizing compiler, 74,480 LOC |
Developer Paul Biggar reported similar results building an optimizing compiler for Darklang over the 2025 Christmas break. Biggar estimated the work would have taken two years of full-time development if implemented manually, according to his blog post. Claude Code wrote every line across 594 commits, producing a compiler with 3,272 tests and 74,480 lines of code.
Productivity Multipliers and Limitations
The Productivity gains extend beyond compiler construction. GitHub reports that developers using Copilot are up to 55% more productive at writing code and report 75% higher job satisfaction, according to GitHub’s official documentation. Microsoft-backed trials show AI assistance leads to approximately a 21% productivity boost in complex knowledge work, per Index.dev research.
However, AI-generated code requires rigorous oversight. A GitClear study found that code copying and pasting has increased since 2022, while refactoring has decreased, leading to higher code churn, and concluded that ‘the rise of AI assistants is strongly correlated with ‘mistake code’ being pushed to the repo’, as reported by Stack Overflow.
Compiler construction traditionally requires deep expertise in parsing theory, abstract syntax trees, intermediate representations, optimization passes, and target architecture code generation. The Scheme language, developed at MIT in 1975, remains a benchmark for compiler implementation due to its minimal syntax but demanding features like proper tail-call optimization and first-class continuations.
Anthropic researcher Nicholas Carlini stated the autonomous compiler experiment both excites him and leaves him ‘feeling uneasy,’ noting he didn’t expect this capability to emerge this early in 2026, according to Medium. Coming from a penetration testing background, Carlini expressed concern about deploying code that no human has personally reviewed.
Shifting Developer Roles
AI agents could handle repetitive tasks, large refactors, testing, and bug fixing, allowing humans to focus on design, architecture, and problem-solving, while raising questions about code quality, trust, and verification, according to Gizmochina.
Engineers seeing massive productivity gains aren’t using magic prompts—they’re using disciplined workflows, with AI coding tools only boosting productivity when used systematically, according to Forge Code.
- AI can now autonomously execute multi-stage compiler tasks that previously required specialized human expertise
- Development time for complex systems compressed from months/years to days/weeks
- Quality control and code review become more critical as AI generates larger codebases
- Developer roles shifting from implementation to architecture, verification, and oversight
What to Watch
The rapid evolution of AI-assisted development raises immediate questions about software supply chain security and verification standards. As AI systems generate increasingly complex codebases with minimal human review, the industry will need robust testing frameworks and verification tools to match the accelerated development pace.
Expect compiler and language implementation projects—once the domain of PhD-level specialists—to become accessible to broader developer populations. The democratization of compiler construction could accelerate language experimentation and domain-specific language development, particularly for WebAssembly targets where Phillips and others have demonstrated viability.
The convergence of AI agents with continuous integration pipelines represents the next frontier. Organizations adopting AI-driven development at scale will need to establish clear governance frameworks around code ownership, liability, and quality assurance as the ratio of AI-generated to human-written code continues to climb.