A System That Lets AI Work Directly With a Real Xbox — Now Open Source

Description and use

XCP implements a development cycle where AI generates code, Xbox executes it via the XVM virtual machine, and the system collects behavioral evidence to correct and evolve the project, ensuring continuous improvement.

Problem

AI-generated code often compiles but behaves incorrectly on the target machine, making it hard to maintain a reliable software lifecycle.

Outcome

XCP demonstrates a full AI‑driven software lifecycle on real Xbox hardware, achieving 28/28 operations for new‑project lifecycle and 14/14 for existing‑source lifecycle (Minilens), with all declared semantic checks passing.

Workflow

• AI agent generates or modifies a project
• XCP builds the project into an XCP project
• XCP prepares the candidate for execution
• Xbox Series X runs the candidate via XVM
• Xbox records structured evidence of execution
• Evidence is returned to XCP
• XCP compares evidence against declared behavioral checks
• Agent corrects or evolves the project based on evidence
• Loop repeats

Evidence

XCP turns one-shot code generation into a continuous software lifecycle: build → execute on Xbox → observe → correct → evolve → repeat

Source

There are currently two main entry points into that lifecycle.

Source

New bounded work can be sent to the execution target without every new behaviour requiring a completely new native application.

Source

Execution is not fidelity.

Source