AI agents build complex software by evolving persistent projects rather than persistent agents

Developing sophisticated software typically requires years of work and continuous knowledge transfer across teams. While most AI coding agents rely on persistent memory or session continuity to maintain progress, a new research paper from arXiv proposes a different approach: keep the software project itself stable and allow individual agents to be swapped out and replaced as needed.

How Genesis Organizes Long-Horizon Development#

The system, called EvoX Genesis, structures code development around what researchers call a "persistent recursive world." Rather than centering continuity on any single AI agent, the framework anchors it on the repository itself. Each local environment is defined by an accepted software version and a repository path. Finite-lived agents propose changes, work distributes across paths through recursive delegation, and only validated changes get incorporated into the permanent version history. This design allows agents to be ephemeral while the codebase accumulates progress over months or years.

Real-World Results#

The team tested Genesis across three major scenarios. In the first case, starting from scratch, they used DeepSeek V4 Flash to autonomously construct a Rust-based C compiler. Over 120 hours of continuous operation, the system produced roughly 250,000 tracked lines of code, cycled through over 1,000 agent episodes, and cost only $44 in API fees. The resulting compiler passed the complete c-testsuite and performed well on LLVM and Csmith test suites. In a parallel experiment with a different model (GLM 5.2), agent replacements occurred repeatedly—the team could swap agents mid-project—yet the compiler retained full test performance throughout.

Genesis also tackled a migration challenge, reimplementing 13 MESA modules (over 100,000 Fortran lines) as a Rust workspace with approximately 90,000 lines. Across six numerical workloads, the Rust versions showed median speedups between 1.55x and 6.87x compared to the originals.

Why It Matters#

The results demonstrate that complex software projects need not be tethered to individual persistent agents. By centering continuity on the codebase rather than the AI, teams can replace, upgrade, or rotate agents without losing momentum. This separation of concerns—stable project, replaceable workers—mirrors how human software teams function and may enable more robust autonomous development systems capable of tackling multi-month engineering challenges.