Sean Goedecke argues that in large software projects, programmers understanding only the code they own is normal and is the best achievable state. Bao Yu extends this to AI agents: they too can never fully understand a codebase in one session, so quality must be backstopped by automated tests, code review, and having agents supply screenshot evidence.
Code Review
代码审查
Human checkpoint reviewing code changes for quality and correctness, now identified as the new bottleneck in AI-assisted develop
Curated Intelligence
The essay argues that software engineering uniquely tolerates the belief that engineers can stop writing code yet keep improving by reviewing and instructing others. Drawing analogies to deliberate practice in music and other fields, the author contends that code review judgment depends on continued hands-on coding and degrades without it.
A Harvard working paper covering 718 firms finds that engineers using AI coding agents produced about 30% more code, yet completed work did not measurably increase and code took 49% longer to pass review. The authors argue the bottleneck has shifted from drafting to human review, and recommend measuring approved output, revision rates and reviewer time rather than raw code volume.
A Tencent agent developer observes that AI-assisted coding has multiplied commit volume dozens of times while controllability declines: lacking the tacit knowledge humans use to trim edge cases during reviews, AI conservatively expands every exception path, inflating coordination state until review degrades into pattern matching. The author argues the bottleneck is convergence, not generation, and calls for explicit pre-generation constraints and human interrupt points.
Baoyu describes porting a team rule—PRs must include screenshots of the changed feature—to AI agents: after finishing a task, the agent self-tests and produces screenshots, often catching and fixing issues along the way. He cites Anthropic's Addy Osmani saying daily merged code changes grew from ~2,500 in May to over 30,000 by September, with humans reviewing AI-submitted evidence such as screen recordings and screenshots rather than reading every line; giving AI a way to verify its own work can improve output quality two to three times.
A large-scale Harvard study of 718 enterprises reveals that while AI coding assistants and agents significantly boost code output (e.g., a 30% increase in lines), this does not translate into proportional gains in completed work. The bottleneck lies in the development lifecycle, where only 30-40% of time is spent on coding, with the rest consumed by planning, testing, and reviews. Data shows AI-generated code increases average merge request review time by 49%, comments by 53%, and revision rates by double, indicating that downstream bottlenecks negate upstream efficiency gains.
A viral meme depicts the 2026 software development norm: two engineers instruct the same LLM to generate and review code, ultimately approving each other's pull requests. This highlights the shift from human coding to human-supervised AI workflows and the resulting ironic feedback loop.
To counter engineer skill atrophy from over-reliance on AI, a 'Brain Sandwich' workflow is proposed: humans first understand the task, then delegate to AI, and finally review and refine the output. This method balances efficiency gains with the preservation of core engineering judgment.
The QGIS project has officially banned AI-assisted code contributions via its Enhancement Proposal (QEP) process to ensure traceability and quality. The policy mandates strict human review for all Pull Requests, preventing unverified automated content from entering the core codebase. This move highlights a broader trend in open-source communities shifting from technical acceptance of LLMs to governance-focused regulation.
A large-scale study across 700 firms reveals that while AI coding agents significantly increase code generation volume, they fail to boost overall software output or reduce employment. The primary bottleneck is human code review, where efficiency gains in the writing phase are entirely absorbed by downstream verification and correction costs.