Tackling Review Fatigue by Document Driven Agentic Coding (chezo.uno)

A senior ML tech lead describes growing “review fatigue” as LLMs accelerate code and doc output without reliably improving quality. He identifies three root problems: “proxy prompting” (where humans simply relay AI-generated content and lose the domain expertise needed to audit it), lack of rigorous self-review (the “eyes glazing over” effect), and agentic LLM behavior that produces big-bang commits and overly complex code. He also warns that LLM reviews aren’t a safe substitute—outputs are inconsistent (“gacha”-like) and incentives differ: generators optimize throughput while reviewers must guarantee quality, creating extra verification work and managerial escalation needs. As a practical counter, he advocates document-driven agentic coding: constrain LLM tasks to ~30-minute human-sized units, enforce a flow of requirements → external design → work plan, and commit those documents (requirements.md, interface.md, plan.md, AGENTS.md) alongside code and acceptance tests. He outlines an agent workflow (Sonnet 4/GitHub Copilot + clear context resets between phases) where each checkbox in plan.md maps to a commit, making intent auditable and reducing pointless low-level reviews. The key implication for the AI/ML community is that workflows and repo-level documentation—not just better models—are essential to preserve human reviewability, enforce granularity, and mitigate fatigue while we wait for more reliable automated review tools.