OurArk: A Framework for Agents That Humans Can Own and Control

Researchers have introduced OurArk, an architectural approach designed to address a fundamental challenge in personalized AI agents: most are configured in ways that give users little visibility or control over what shapes the agent's future decisions and behaviors.

The Core Concept#

At the heart of OurArk is the notion of an agent-owned software body—a persistent, version-controlled artifact that humans can inspect, audit, and modify. Rather than treating agents as opaque black boxes, this body is a tangible digital asset under human custody. It encompasses the agent's behavior-defining code, prompts, embedded tools, learned skills, operational policies, and validation mechanisms. Importantly, private information like memories and credentials are kept as instance state separate from the body, while AI model inference itself is treated as a replaceable external service rather than something fixed in stone.

Self-Evolution and Descent#

The framework introduces two key mechanisms for change. Self-evolution allows an agent to propose modifications to its own software body, but these changes are isolated as candidates that must be validated, reviewed, and approved by a human before being merged into the live version. This creates a structured path for human-agent collaboration in developing the agent's capabilities.

A second mechanism, called descent, enables the creation of new, independently versioned agents derived from an existing parent. Each descendant has its own distinct identity, mission, and history—with a fresh boundary around its private state. Once created, descendants can themselves spawn further descendants. After diverging, child agents can inspect both direct-parent modifications and skills from peer agents, then selectively adapt those changes locally.

Implementation and Demonstration#

The team implemented OurArk in the open-source Genesis creation engine and built a reference agent called Enoch. Their proof-of-concept includes a four-agent lineage across three descent events, complete with executable regression tests. This demonstration showcases recursive agent creation, inherited validation rules that persist across lineages, isolated code changes, human-controlled review workflows, and recovery from failed updates.

Why It Matters#