Developer Builds Multi-Agent AI Code Review System That Negotiates Disagreements

A developer created ShiftLeft Society, a multi-agent DevSecOps code review system, for the Qwen Cloud Global AI Hackathon 2026, addressing a structural flaw in existing AI review tools. The core problem identified was that current tools, including newer multi-agent platforms, lack a mechanism to resolve disagreements between agents, leaving that burden to developers. The system features two AI reviewers that negotiate conflicts using a budget-based cost structure, where agents choose categorical positions — Defend, Partial, or Concede — while deterministic Python handles all numerical calculations to ensure auditable and reproducible verdicts. Each agent also maintains a running track record across reviews, using Bayesian smoothing to adjust its starting negotiation budget over time, allowing the system to learn which agent's judgment is more reliable. In benchmark testing, the tribunal achieved 95% accuracy across 40 cases, compared to 82.5% for a single-agent baseline, with most gains coming from reduced false positives on safe code.