Coding Agent Reliability EXPLODES When They Argue (New Adversarial Dev Technique)
Your AI coding agents fail on complex tasks not because the model is insufficient, but because you aren't separating the planning, building, and evaluation into distinct roles. The building and evaluation agents essentially have to fight each other to get the best results. Anthropic just published a deep engineering article breaking down exactly how they solved this for multi-hour coding sessions with a GAN-inspired three-agent architecture, and the pattern is something you're going to want to try on your own codebases! But I didn't just want to explain the article. I built my own version of this three-agent harness using the Claude Agent SDK and the Codex SDK, and I'm giving it away. ~~~~~~~~~~~~~~~~~~~~~~~~~~ - Want to learn AI engineering? Scrimba's AI Engineer Path is hands-on, project-based, and very effective: https://scrimba.com/learn/aiengineer?via=colemedin - All Scrimba coding courses (many free): https://scrimba.com/courses?via=colemedin ~~~~~~~~~~~~~~~~~~~~~~~~~~ - If you're interested in building your own AI second brain to save yourself hours every week, check out the Dynamous community and upcoming 4 hour second brain workshop: https://dynamous.ai/second-brain-bootcamp ~~~~~~~~~~~~~~~~~~~~~~~~~~ - Adversarial Dev Harness (my implementation - Claude Agent SDK + Codex SDK): https://github.com/coleam00/adversarial-dev - Anthropic: Harness Design for Long-Running Application Development: https://www.anthropic.com/engineering/harness-design-long-running-apps - Anthropic: Effective Harnesses for Long-Running Agents (companion piece): https://www.anthropic.com/engineering/effective-agent-harnesses ~~~~~~~~~~~~~~~~~~~~~~~~~~ 0:00 AI's Sycophancy Problem 1:02 Adversarial Dev Solution 2:31 GAN-Inspired AI Harnesses 4:23 Harnesses Build Amazing Apps 6:49 Scrimba 8:07 Harness System Architecture 10:39 Sprint Negotiation & Evaluation 12:48 Get Started With Harness 16:01 Token Cost vs. Value ~~~~~~~~~~~~~~~~~~~~~~~~~~ Join me as I push the limits of what is possible with AI. I'll be uploading videos weekly - at least every Wednesday at 7:00 PM CDT!
