OurArk lets AI agents own, fork, and evolve their code bodies

Personal AI agents have a transparency problem: you can configure them, but you can't control the code that shapes their future behavior. OurArk, a new architecture from Roy Zhao (University of Washington) and Zhenyu Zhao, flips that by giving each agent an identity-bearing, versioned software body—wrapping behavior code, prompts, tools, skills, policies, and tests in a single artifact the user actually owns. Memory and credentials stay private instance state, and model inference is treated as a swappable external service, so the body is portable across models.

The system introduces two powerful mechanisms: self-evolution, where the agent proposes isolated candidate changes to its own code that must pass validation and human review before merging, and recursive descent, which forks an independently versioned descendant with a fresh identity, mission, and private-state boundary. Compatible descendants can keep forking, and after divergence, parent changes and peer skills remain inspectable for selective adaptation. The authors implemented both in the open-source Genesis creation engine and the Enoch reference agent. Their experiments—a four-agent, three-descent linear lineage plus executable regression tests—demonstrate recursive creation, inherited validation contracts, isolated changes, human-controlled merging, and recovery from failed updates. It's a concrete substrate for agents people can genuinely govern and specialize over time.