Adversarial Dev Harness (my implementation - Claude Agent SDK + Codex SDK)

This is my implementation of the three-agent adversarial harness: planner → generator → evaluator, with sprint negotiation and scoring thresholds. I built it because the Anthropic article didn’t ship code, and I wanted something you can clone and run in minutes to see the reliability jump for yourself.

Pros

+Implements planner/generator/evaluator loop with sprint negotiation
+Supports Claude and Codex versions (and you can mix implementer vs evaluator)
+Designed for quick local experimentation and proof-of-concepts

Cons

-Higher token usage due to iterative evaluation loops
-Not a guarantee of production-ready code; it’s a harness for faster POCs

Featured in this video

Coding Agent Reliability EXPLODES When They Argue (New Adversarial Dev Technique)

9K views · 2026-03-30 16:30:05

Adversarial Dev Harness (my implementation - Claude Agent SDK + Codex SDK)

Github Available on Github

Buy on Github

External purchase

Highlights

+Implements planner/generator/evaluator loop with sprint negotiation
+Supports Claude and Codex versions (and you can mix implementer vs evaluator)
+Designed for quick local experimentation and proof-of-concepts

Recommended by

Cole Medin