Checking agent code is the slow part.

What happens to code review as AI use rises?

In Faros AI's 2026 telemetry (22,000 developers, vendor data), teams' highest-AI-use quarters showed 66.2% more epics per developer than their lowest, but median time in PR review was 441.5% higher and deployments per week 11.7% lower. This is a correlation, not a measured cause.

Where the code review bottleneck is

• 96%

  Of developers don't fully trust that AI code is functionally correct. Only 48% say they always check it.

  Sonar, January 2026 · 1,100+ developers · vendor

• +199.6%

  Average increase in time spent in PR review; +441.5% at the median.

  Faros telemetry, 2026 · 22,000 developers · vendor

• 66%

  Of developers cite "AI solutions that are almost right, but not quite."

  Stack Overflow Developer Survey 2025 · 49,020 respondents

Where the time goes

1. Agent

   Write code

   PRs are 51.3% larger and edit 59.7% more files.

   Faros, 2026

2. Human

   Wait for review

   Median time to first review: +156.6%.

   Faros, 2026

3. Human

   In review

   Median time in review: +441.5%.

   Faros, 2026

4. Outcome

   Merged unreviewed

   PRs merged without any review: +31.3%.

   Faros, 2026

5. Outcome

   Deploys

   Deployments per week: −11.7%, in the ~10% of teams that measure it.

   Faros, 2026

Ideatives’ view: the cost has moved from writing code to judging it, including deciding what to leave alone. No source measures this directly.

What to do

1. Write down the problem and the "do not touch" list before an agent starts.
2. Cap PR size. Require a named human reviewer for agent-written changes.
3. Track review time and rework instead of lines produced.

Questions

Sources

• JetBrains Research, How Much Code Do Developers Really Let Agents Write?, August 2026. 15,000+ respondents. Self-report; averages bracket midpoints, and its three shares sum to about 112%. JetBrains sells AI coding tools.
• Sonar, State of Code press release and blog, January 8, 2026. 1,100+ respondents. Puts AI-generated or assisted code at 42%, a broader measure than JetBrains'. Sonar sells code verification.
• Faros, AI Engineering Report 2026. 22,000 developers, 4,000 teams, analysis as of March 2026. Telemetry; compares each team's own quarters. Faros sells engineering analytics.
• Stack Overflow Developer Survey 2025, AI section. 49,020 respondents. Only 14.1% used AI agents daily at work.
• Every source here sells something related to AI coding. No independent or academic source was found.