Trade line-by-line code review for AI-agent debate and shared decision-making

Six months of watching teams implement AI-assisted code workflows has revealed a fundamental mistake: automating human judgment while expecting humans to meaningfully review AI's output. The result is "review theater"—a human skimming a PR while two AI systems have already negotiated the actual changes.

That's not how to evolve code review. Instead, a five-layer approach relocates each task to where it belongs: letting AI agents argue before the PR exists, capturing decision-making as work happens, codifying repeated feedback into automated checks, preserving human debate about architectural choices, and clarifying who owns the standards that govern code.

Why code review matters (and it's not bugs)#

Research consistently shows developers cite defect-finding as their main reason for reviewing code—yet actual review comments focus far more on knowledge transfer, architecture alignment, and team-wide understanding. A 2013 Microsoft study and a follow-up at Google both found the same pattern: bug-catching was a small fraction of what actually happened in review threads.

The real cost that hasn't dropped is understanding. AI has made writing code cheap; comprehending the system's rationale, intent, and constraints remains expensive. Reviewing code was how teams built shared mental models. Automating it away without replacement deletes institutional knowledge.

Faros AI's 2026 report covering 22,000 developers found PRs merged with zero human review jumped 31.3%, while incidents per PR rose 242.7% and bugs per developer climbed 54%. The machine was never going to read those changes.

The five layers: what AI does, what humans decide#

Argue: Let AI agents disagree before the PR is created. If one model writes code and another reviews it, that negotiation should happen offline, surfacing a record of proposals, rejections, and reasoning. Humans then debate what the machines couldn't settle—the actual architectural or design choices—instead of drowning in diff context. Tools like PR-Agent, Aider, and AutoGen already support multi-agent workflows.

Capture: Record intent and acceptance criteria as they emerge during development. Every time an engineer corrects or questions an AI agent, a decision is made. Publish those decisions on the PR so reviewers see what was actually settled, not just a finished diff. Specs written upfront drift; living documentation of choice captures the real story.