gan-style-harness

A GAN-inspired multi-agent architecture improves app development by separating generation from evaluation, creating an adversarial feedback loop. This avoids self-critique bias and forces iterative refinement until the application meets rubric-driven production standards.

You can build an application from a concise prompt by running a sprint-style multi-agent workflow. A Planner drafts the scope, a Generator implements the app, and an Evaluator runs Playwright-based live testing and rubric scoring across iterative cycles.

Yes, Playwright-based live testing is integrated directly into the evaluation phase of the generation loop. The Evaluator agent executes live tests to validate the generated application's functionality against the configured rubric criteria.

You can configure max iterations and scoring thresholds using environment variables. These settings control the adversarial loop's stopping criteria, ensuring the Generator and Evaluator cycle until the application meets the defined rubric-driven score threshold.

The best way to avoid self-critique bias is using an adversarial architecture that separates the Generator and Evaluator agents. By isolating evaluation into a dedicated agent with rubric-driven scoring, you prevent the generation logic from justifying its own flaws.