Leverage Code Review for Sustainable AI Coding Development

As the volume of AI code outstrips the ability of developers to review it, the role of developer peer review is gaining critical importance in the software development lifecycle.1

In the pre-AI coding era, peer review served as a collaborative dialogue that focused on increasing code quality and reliability, identifying defects and bugs, sharing knowledge across teams, and enforcing coding standards and consistency. Today, code review is evolving into an indispensable risk management and verification tool to ensure that the output of AI coding tools is sustainable, functional, and secure.

The evolution of peer review—also known as code review—is designed to prevent security vulnerabilities, code churn, design flaws, and technical debt caused by AI coding tools. Over time, risks within the AI coding process can increase costs, suck up senior developer time, and bog down product launches.

“I think of code review as really looking at code for whatever it is that you’re most concerned about,” said Jason Cohen, founder of SmartBear, a coding application integrity tool, and author of Hidden Multipliers: Jump Start Growth with Your Existing Budget and Team. “Maybe you’re looking at security or accuracy or refactoring—whatever your goals are for your code. I don’t see any difference between code review and peer review.”

As AI coding tools continue to evolve towards even more autonomous systems designed to perform with fewer built-in human checkpoints, researchers say organizations and developers need to develop and maintain rigorous review processes to fully leverage the capabilities of AI coding tools.2

How Peer Review is Changing

Before AI coding, developers wrote code and other developers reviewed it. The review process varied historically, occurring in some cases through formal, checklist-based inspections, but more commonly through a conversational process between individual developers or a team of developers focused on catching defects, improving the code’s long-term maintainability, and spreading knowledge of the codebase across a team.3

Today, with AI coding tools and, increasingly, AI agents doing the coding, the role of developers has shifted.

“Before, one person or a team would generate code; now, AI agents create large code changes within minutes and sometimes even within seconds,” said Michaela Greiler, a software engineering researcher, consultant, and trainer.

“As the changes increase, it’s getting harder and harder for human reviewers to look at the code and give meaningful feedback,” she continued. “There are larger code changes, a higher volume of code, which means code review is shifting from a team practice to an individual developer practice.”

For Nick Chapsas, founder at Dometrain, a software engineering training company, peer review has moved beyond a process of checking syntax and style. “Reviewers need to ask whether the code solves the right problem, whether the APIs and configuration are real, whether the tests reflect the requirement, and whether the change introduces hidden costs such as extra network calls, retries, logging, dependencies, or infrastructure usage,” he said.

Tolga Tarhan, founder of Atomic Gravity, an AI software engineering firm, noted that the entire shape of code review has changed. “It used to be a gate at the end of the development process, where a senior engineer reviewed a pull request after work was essentially complete,” he said. “Today, review happens continuously throughout the development process.”

How to Use Peer Review in AI Coding

To optimize the role of peer review in AI and AI agent coding, an organization should first determine the role of peer review in its coding development process.

“The most important thing we’re doing in engineering right now is creating the space to earnestly talk about why we do code review at all, and reaching an explicit, shared agreement about what ‘good’ looks like,” said Josh Parsons, senior software engineer II at Honeycomb, a software observability platform.

“Establishing that foundation will set the table for a much harder conversation as pull request volumes increase: what would it take for some classes of code to be merged without a human ever reading them?” he asked. “You can’t answer any question of code review automation honestly until you can agree at the organizational and team levels what your cultural work norms and values are.”

At Threada, an AI software platform, AI assists in codebase development, while developers own the architecture, acceptance criteria, and release decisions. “Reviewers get little value from re-reading hundreds of plausible-looking lines,” said founder Nick Balnaves. “They focus on assumptions, trust boundaries, and evidence.”

“Our workflow uses three layers: automated checks selected from the files affected, a fresh adversarial review whose brief is to find bugs, security regressions, and false assumptions, and an end-to-end validation of the real workflow,” he added.

When done correctly, code review ensures that developers understand the code that an AI coder writes before it goes into production, said Snigdha Alathur, a data and analytics engineering lead at Spring Point Technologies, an IT services and consulting firm. “Some teams now require a human or an AI-assisted but human-written test case to write the tests instead of the AI,” she said. “Writing the test forces you to actually understand the code.”

Cohen said an important function of code review for AI coding tools and AI agents is to encode knowledge around the bigger patterns or ideas in coding that the AI platform didn’t understand. “Even when you encode that knowledge again, it won’t always get that right, but over time it will get better and as it improves, the investment of your time is worthwhile,” he continued. “Just like you invest in a junior developer, you invest in improving the AI coding platform because that extra bit of time means that it won’t always make the same mistake over and over again.”

The consequences of a lack of review in AI coding are stark, according to Sergey Matikaynen, co-founder and CTO of GoGloby, an AI software development implementation company. “Pull request review times go up, and so do failure rates,” he said. “Going fast without reliability is just running into trouble. You’re not actually shipping faster, you’re just delaying the needed review, which often results in incidents, rollbacks, or senior engineers spending their weekend untangling the mess from someone else’s AI.”

To benefit from AI’s acceleration of the software development lifecycle, sources said organizations and teams must implement and enforce process changes.

“For organizations as a whole, the question becomes: how do we make sure that everyone has access to new ways of working [in peer review],” said Greiler. “We also need to ensure transparency in communication so that developers are showcasing their good practices or use cases and the problems they run into in terms of how they are using AI and AI agents in coding.”

References

1. The Acceleration Whiplash: AI Engineering Report, Faros (2026). https://pages.faros.ai/hubfs/AI_Engineering_Report_2026_The_Acceleration_Whiplash_Faros.pdf

2. Governed AI-Assisted Engineering: Graduated Human Oversight for Agentic Code Generation in Regulated Domains, arXiv (July 4, 2026). https://arxiv.org/pdf/2606.22484

3. Monperrus, M. The End of Code Review: Coding Agents Supersede Human Inspection, arXiv (June 11, 2026). https://arxiv.org/pdf/2606.13175

Amy Buttell is a Silver Spring, MD-based technology, legal, and business journalist, content creator, writer, and ghostwriter.

Submit an Article to CACM

CACM welcomes unsolicited submissions on topics of relevance and value to the computing community.

You Just Read