We open-sourced Myli: a harness for AI design agents (github.com)

Myli, a new open-source Python harness for AI design agents, has been released to facilitate the modification of JSON design documents without persisting changes to application state. Operating independently of various frameworks and libraries, Myli allows developers to implement agents that propose RFC 6902 changes through a robust API. The system promotes safe design through rigorous candidate validation against predefined schemas, ensuring that any modifications are carefully controlled and audited. This makes Myli particularly significant for AI/ML applications where designing and maintaining complex data schemas is critical. The harness supports a wide range of features, including parallel execution of tasks, detailed outcome recording, and customizable middleware for auditing and governance. With built-in mechanisms for handling rendering and visual feedback, Myli enables applications to adapt designs dynamically while applying strict limits on model operation and effectiveness. By enhancing how AI agents interact with JSON documents, Myli not only improves operational efficiency but also encourages the adoption of best practices in data handling and design within the AI/ML community.