Adversarial Dev

Adversarial Dev

A GAN-inspired three-agent harness that separates **planning**, **building**, and **evaluation** into distinct AI agents with distinct contexts. The evaluator's job is to **break** what the generator builds -- creating adversarial tension that drives quality far beyond what a single agent can achieve. Built with both the **Claude Agent SDK** and **Codex SDK** so you can run the same architecture on either platform.

Based on Anthropic's engineering article: [Harness Design for Long-Running Application Development](https://www.anthropic.com/engineering/harness-design-long-running-apps).

What This Demonstrates

Most AI coding agents fail on complex tasks not because the model is bad, but because nobody separated the work into specialized roles. A single agent that plans, builds, and evaluates its own work will reliably praise its own mediocre output. This is called **self-evaluation bias**, and it's the quiet killer of ambitious AI coding projects.

This project implements the fix: three agents, each with a focused job and its own context window.

Agent	Role	Analogy
Planner	Expands a short prompt into a full product spec with sprints	Product manager
Generator	Builds one feature at a time, commits to git	Software engineer
Evaluator	Actively tries to break what the generator built, scores ruthlessly	Adversarial QA

The evaluator doesn't just review code -- it's an adversary. It runs the application, probes for failures, tests edge cases the generator didn't think of, and scores each criterion on a 1-10 scale with a hard pass threshold. If any criterion fails, the sprint goes back to the generator with detailed, unforgiving feedback. The generator has to fight its way past the evaluator to advance. This adversarial pressure is what turns AI-generated code from "looks right" into "actually works."

Quick Start

Prerequisites

• Bun runtime installed
• Cla