Research & Papers

EvoX Genesis builds Rust C compiler with $44, 120-hour autonomous run

Persistent recursive worlds let finite-lived AI agents build 250k-line compilers autonomously

Deep Dive

EvoX Genesis, introduced in a new arXiv paper, challenges the assumption that long-horizon software development requires persistent, long-lived AI agents. Instead, Genesis keeps the software project persistent as a 'recursive world'—each local world is anchored by an accepted version and repository path, with finite-lived agents proposing changes. Recursive delegation lets tasks move across paths, and only accepted changes advance the version history. This design decouples project continuity from agent lifespan, enabling autonomous evolution over timescales that exceed any single agent's session.

In evaluation, Genesis demonstrated remarkable efficiency and scale. Using DeepSeek V4 Flash, it built a Rust-based C compiler from scratch—about 250,000 tracked lines—in over 120 hours of runtime, spanning more than 1,000 agent episodes and costing just $44 in model-token charges. The compiler passed the complete c-testsuite and most LLVM and Csmith tests, indicating production-grade quality. Separately, Genesis continued development in a compiler world generated with GLM 5.2 even after repeated agent replacement, retaining full test performance. In a redevelopment scenario, it migrated 13 MESA modules with over 100k Fortran lines into a Rust workspace of nearly 90k lines, achieving median speedups of 1.55-6.87x across six numerical workloads. These results show that persistent projects, not persistent agents, can drive long-horizon autonomous software evolution.

Key Points
  • Genesis used DeepSeek V4 Flash to build a Rust C compiler with ~250k tracked lines over 120 hours, costing only $44 in token charges
  • The compiler passed the complete c-testsuite plus most LLVM and Csmith tests, demonstrating production-quality output
  • It reimplemented 13 MESA Fortran modules (100k+ lines) as a Rust workspace with 1.55-6.87x median speedups across six numerical workloads

Why It Matters

Enables low-cost, autonomous long-horizon software projects by making code persistent, not agents, which could reshape AI-driven development pipelines.

📬 Get the top 10 AI stories daily