# Leverage Code Review for Sustainable AI Coding Development

A Communications of the ACM news article by Amy Buttell, published 4 September 2026. Its thesis is that as AI output outstrips developers' ability to review it, peer review turns from a collaborative quality practice into a risk-management and verification step, the thing that keeps AI-written code sustainable, functional and secure. The piece is built from interviews with founders, consultants and engineers, and has no data of its own. It lists three references (the Faros "Acceleration Whiplash" AI engineering report, an arXiv paper on graduated human oversight for agentic code generation in regulated domains, and Martin Monperrus's "The End of Code Review: Coding Agents Supersede Human Inspection") without quoting or summarising any of them. The last title argues roughly the opposite of the article's headline.

## How the sources describe the change

Michaela Greiler, a software engineering researcher and trainer, says agents create large changes in minutes or seconds, making meaningful human feedback harder, and that review is shifting from a team practice to an individual developer's practice. Tolga Tarhan of Atomic Gravity says review used to be a gate at the end, where a senior engineer read a finished pull request, and now happens continuously during development. Nick Chapsas of Dometrain lists what reviewers should now ask: whether the code solves the right problem, whether its APIs and configuration exist, whether the tests reflect the requirement, and whether the change adds hidden costs such as extra network calls, retries, logging, dependencies or infrastructure use.

## How they say to do it

Josh Parsons of Honeycomb puts agreement first: teams should talk openly about why they review at all and what "good" means, because that sets up the harder question of what it would take for some classes of code to merge without a human ever reading them.

Nick Balnaves, founder of Threada, gives the most concrete workflow. AI helps build the codebase while developers own architecture, acceptance criteria and release decisions, and reviewers get little value from re-reading hundreds of plausible-looking lines, so they look at assumptions, trust boundaries and evidence. The pipeline has three layers: automated checks selected by which files a change touches, a fresh adversarial review whose brief is to find bugs, security regressions and false assumptions, and end-to-end validation of the real workflow.

Snigdha Alathur of Spring Point Technologies describes teams that require the tests to be written by a human, or AI-assisted but human-written, because writing a test forces you to understand the code. Jason Cohen, founder of SmartBear, sees review as the place to encode the larger patterns the AI missed, an investment comparable to training a junior developer so the same mistake stops recurring. Sergey Matikaynen of GoGloby describes skipping review as deferring it: review times and failure rates climb, and the cost comes back as incidents, rollbacks and senior engineers' weekends.

## Where it sits

This is the "review matters more" side of the argument. The opposite side is [[laycock-review-by-exception]], which says most of review's jobs should move earlier and human review should become the exception. Parsons's question about merging without human reading points in Laycock's direction. Balnaves's reason for not re-reading plausible lines is the throughput problem in [[code-review-throughput-limits]] and [[reviewing-ai-code]], and his "fresh adversarial review" is the practice examined in [[cross-model-code-review]] and the adversarial subagents in [[llm-critics-are-right-use-anyway]]. Alathur's human-written tests echo [[testing-heavy-no-review-workflow]], with the twist that the test's value is the understanding it forces, not only the defects it catches.

## Cross-references

- [[agent-principal-agent-problem]] — why review's old effort signal failed once agents wrote the code
- [[skill-atrophy-supervision-paradox]] — the risk under Cohen's and Alathur's advice: supervising agents erodes the skills supervision needs
- [[google-code-review-looking-for]] — the pre-agent baseline of what a reviewer checks