Software engineers using coding agents in 2026 are generating more code than at any point in the industry’s history, and shipping less of it intact. That is the pattern developer advocate Addy Osmani drew out in an analysis posted on X this week, synthesizing two datasets that together describe the same production failure from different angles.
The larger dataset comes from Faros AI, a workforce analytics company whose study covered 22,000 developers. The numbers are striking on their own: code churn (code written and then deleted or rewritten within 21 days) rose 861%. Per-developer defect rates climbed from 9% to 54%. Review duration stretched 441%. Zero-review merges, the most direct measure of quality gates collapsing, rose 31%. These figures are not describing a marginal change in team hygiene. They describe a system under structural stress.
The second source, GitClear, frames the productivity gap in a single ratio: teams using coding agents achieved roughly four times the raw output, but captured only about 12% more delivered value. The gap between those two numbers is not a rounding error. It is the cost of reviewing, debugging, and discarding code that was generated quickly but not generated correctly.
Osmani’s framing is that the hard part of engineering has migrated. Writing code used to be where skill accumulated and where bottlenecks formed. Now code arrives in quantity. The bottleneck is deciding which of it to trust, at what point in the stack, and with how much verification. That is a review problem, and the data suggests most organizations have not restructured their processes to treat it as one.
One structural skepticism worth naming: the Faros AI and GitClear studies measure different things and were conducted independently. Faros AI is measuring developer activity signals across a large enterprise sample; GitClear is measuring commit-level patterns. Neither study was designed as a controlled experiment comparing AI-assisted teams against matched controls with identical project complexity. The 861% churn figure and the 4x output number are directionally consistent, but they do not come from the same cohort, and the magnitude should be treated as a range, not a precise benchmark.
That caveat does not weaken the core finding. It sharpens it: across two methodologically distinct datasets, the signal pointing toward a review crisis is present and large.
For engineering leaders, the 90-day decision is not about whether to adopt coding agents. Most teams already have. The decision is whether the review process has been redesigned to match the new input volume. A 441% increase in review duration with a 31% rise in zero-review merges means reviewers are drowning and some are giving up. The remediation path is not more reviewers. It is scoped review gates: automated checks that catch the highest-frequency defect categories before a human sees the diff, smaller pull request norms that reduce cognitive load per review, and explicit team policy on which layers of the stack require human sign-off regardless of agent confidence.
Teams that do not make that adjustment are trading a short-term output gain for a medium-term debt accumulation. The Faros AI and GitClear data together suggest that debt is already accruing faster than it is being paid down.
Engineering leaders who have not audited their zero-review merge rate in the last 60 days should treat that number as the first metric to pull before deciding whether their current review process is adequate for an agent-assisted team.
Analysis by developer advocate Addy Osmani, published on X on June 16, 2026, synthesizing data from Faros AI’s 22,000-developer study and GitClear’s commit-pattern research.
