AI Coding Paradox: 30% More Code, Zero Output Gain — Why Trust Is the Real Bottleneck

Study Scope and Methodology

Harvard researchers Fiona Chen and James Stratton analyzed aggregated engineering data from Jellyfish, covering 300 million work events (commits, pull requests) and issue-tracking data from over 700 software companies employing more than 700,000 engineers between 2021 and March 2026. They combined direct AI-usage measurements with GitHub activity analysis to pinpoint when each organization introduced AI coding assistants (which autocomplete code for humans) or AI coding agents (which autonomously write and submit code from prompts). A difference-in-differences regression compared key variables before and after tool adoption across organizations at different times.

Raw Code Output Rises, Software Output Does Not

After introducing AI coding agents, companies saw a 30% increase in total lines of code generated, a 20% rise in commits, and a 23% increase in pull requests. However, Jira-tracked issue and epic resolution rates — the proxy for delivered software features — showed no statistically significant change. The extra code did not translate into higher feature throughput.

Code Review Becomes the Bottleneck

The divergence is explained by the review process. Average time from pull-request submission to merge — the "review cycle" — inflated by 49% after AI agent adoption. The proportion of engineers performing reviews grew by 14%. Cross-referencing Jellyfish active-user counts with LinkedIn headcounts ruled out hiring changes as a cause. Although 80% of surveyed firms used some form of AI code review by March 2026, AI agents contributed only 23.3% of all review comments and 10.8% of pull-request reviews, leaving the vast majority of verification work to humans.

Trust, Not Velocity, Is the Core Problem

AI agents can produce initial solutions in minutes and raise output volume, but the surge in changes (code edits, service refactors, test additions) outpaces human review capacity. The critical challenge for software organizations is not writing more code but establishing confidence in code that reaches production. A merged pull request does not guarantee reliability: code may compile and pass tests yet violate architectural conventions or introduce security vulnerabilities.

A Framework for Code Confidence

The article proposes embedding code reliability into the software delivery system through four pillars:

Intent — Engineers and product managers must define business logic, security expectations, design patterns, and performance requirements before the agent starts work.

Guiderails — Intent is translated into operational boundaries: allowed repositories, dependencies, data, operations, and human approval gates.

Verification — Independent checks of every change for correctness, reliability, security, maintainability, and organizational standards. The same probabilistic system that produced the change should not be the sole judge of its safety.

Evidence — A durable record of applied standards, executed checks, findings, and resolutions, making quality measurable and auditable.

This approach helps teams improve practices over time and provides a trustworthy foundation for rapid development. The most effective vision for agentic software development integrates people, agents, and verification systems so each plays to its strengths.

Conclusion

As of the March 2026 data cutoff, 95% of studied companies had adopted AI coding agents, yet many are still experimenting with when and how to deploy them effectively. For now, AI coding resembles a double-edged sword: speed gains are offset by increased manual review effort, raising the question of whether the investment in AI agents is worthwhile for most organizations.