AI Papers Reader

Personalized digests of latest AI research

View on GitHub

When AI Agents Expire, the Software Project Carries On

Long-lived software has a continuity problem: the coding agent that starts a project rarely survives to maintain it. Most AI coding systems try to fix this by giving agents longer memory, bigger context windows, or persistent managers. Researchers at The Hong Kong Polytechnic University describe a different approach in a new preprint: keep the project, not the agent, alive.

The EvoX Genesis system represents software as a “persistent recursive world.” Each finite-lived agent is dropped into a local world defined by two coordinates: an accepted version of the codebase and a repository-relative path. The agent works on a bounded task, proposes a change, and terminates. Its private conversation dies with it. What survives is the accepted version history and path-specific context files. Managers can delegate subtasks to child agents at other paths without changing the accepted version; a parent must accept a returned result before it becomes part of the persistent history. Rejected changes leave the project untouched.

The approach produced three demonstrations. In a “formation” test, the system used DeepSeek V4 Flash to build a Rust-based C compiler from a repository containing only a .gitignore file and a config file. Over 123.4 hours, 1,019 archived agent episodes accumulated 248,989 physical lines of tracked code—including comments and blank lines—at a model-token cost of US$44.38. The compiler passed 220/220 c-testsuite tests, 93/93 executed Csmith programs, and 32/36 evaluated LLVM test cases; the researchers report the suites separately because their denominators differ. No single episode spanned the project. The authors note that later integration could expose problems in components that had seemed complete locally, and that repairs then occurred with the rest of the project in place.

In a continuation experiment, two branches started from the same compiler previously generated by GLM 5.2. One branch continued with GLM 5.2, passing 1,445/1,448 cases in its snapshot’s LLVM test set; the other switched to DeepSeek V4 Flash and passed 1,820/1,820 on its own snapshot’s set. Because the lists differ, the authors say this is not a head-to-head comparison. In a redevelopment test, 13 MESA astrophysics modules containing 139,414 physical Fortran lines were reimplemented as a Rust workspace with 89,946 physical Rust lines. Across six numerical workloads, median runtimes were faster in Rust by 1.55× to 6.87×; two workloads were bit-exact, and relative checksum differences for the other four ranged from 5.1×10⁻Âč⁔ to 3.1×10⁻âč. The timing comparisons apply only to the reported builds and host.

The authors are careful about what this shows. The runs are single observations, not controlled comparisons; recursive delegation was used extensively, but the study did not test whether flat alternatives would do as well. “Autonomous” is also bounded—humans supplied objectives, tools, and validation limits. The reported dollar figures are token charges only. Still, the work suggests a broader hypothesis: a persistent project state can carry software development across changing episodes of intelligence, even when no single AI coder remains.