I Built a Virtual Machine Inside the Xbox Sandbox. Then I Let AI Agents Build on Top of It.

The XCP framework transforms a one-shot code generation process into a continuous software lifecycle, enabling AI agents to generate software and then iterate on it. The innovation lies in the fact that the generated code can still be incorrect even after successful compilation, or a generated project can exhibit unexpected behavior when launched.

XCP aims to resolve these issues by having the software project continue beyond generation, allowing it to change, run on a real Xbox Series X, generate evidence, and be corrected and evolved into a new version. This loop ensures that the software project survives the generation process and retains its identity, with executions and evidence tied to specific versions.

The framework's core idea is that a loop should exist between AI agents and development, rather than the traditional AI agent → source files → build → done approach. This allows for a more dynamic and interactive development process, where the Xbox console becomes an integral part of the development loop. XCP currently enters the lifecycle in two ways: starting from an idea or from existing authorized software.

The first case involved a public creation program where an AI-driven, human-directed workflow completed all lifecycle operations across an interactive project and a utility, demonstrating that AI can produce code and continue through the lifecycle. The second case used an existing Godot 2 project, Minilens, and successfully completed all lifecycle operations, including semantic acceptance checks on the Xbox.

The framework's virtual machine (XVM) allows for programmable execution on a retail Xbox Series console while staying within the public application environment. XVM provides a programmable target for XCP and ensures that new bounded work can be submitted to the execution target without every new behavior becoming a completely new native application.