How we review 4x more AI code without reading every line
B · Useful, limited Aviator's CEO describes a concrete review process with before/after numbers from his own team in a sponsored New Stack post; self-reported vendor figures with no defect-escape data.
Summary based on the full text·Original: about 10 minutes to read
In short
A New Stack post (7 November 2026, sponsored by Aviator) by Ankit Jain, Aviator's co-founder and CEO. As AI agents wrote most of the code, his team stopped reviewing diffs line by line and started reviewing five things: intent, acceptance criteria, invariants, decisions and evidence. A person still signs off every PR. The key move was mining past review comments into rules that machines check on every change.
Key numbers
- 4xPRs merged per week with the same nine engineers (about 100 to more than 400)
- 2.1 days to 3.4 hoursmedian time to merge
- 40%drop in reverts per 100 PRs
- 11,400 to 64review comments mined from a year of history, reduced to 64 approved invariants (212 candidates)
- 6.2 to 1.3human review threads per PR
As reported by the source.
English·Web page·opens thenewstack.io
What it reports
- Weekly merged PRs went from about 100 to more than 400; the median AI-generated PR reached 640 lines.
- What failed: a generic AI reviewer (16 of 18 threads on one PR were its nits), a four-hour review SLA (faster, emptier approvals), and a convention-tuned self-review skill that still read the diff after the fact.
- Before the PR: the review agent, on a different model, cross-examines the author's agent and records disagreements.
- Acceptance criteria are captured from each correction or answer an engineer gives the agent during the session.
- A year of review history: 11,400 comments became 212 candidate rules and 64 approved invariants, each with scope, verification method, provenance and owner; 19 compile to ast-grep rules running in 0.3 seconds.
- Evidence: an agent drives a preview environment and records screenshots, requests and logs for each criterion.
- Results: 4x merged PRs with the same nine engineers, median time to merge 2.1 days to 3.4 hours, human threads per PR 6.2 to 1.3, reverts per 100 PRs down 40%. Transition: double work in month one.
How they got it
The author's account of his own team's metrics; no method, period length or statistical detail.
Limits
- Vendor CEO describing a process his product sells; sponsored content.
- Reverts are a weak quality proxy; no data on defects reaching production.
- One small team; results may not transfer to large or regulated codebases.
Why it matters for companies
When agents produce code faster than people can read it, review capacity becomes the throughput limit. Converting repeated reviewer comments into owned, testable rules frees people for the decisions a diff never shows. The mining exercise (comments to rules to deterministic checks) is something any engineering organisation can run on its own review history without buying a tool.
Summary based on the full text. Read the original for the full argument and data.Read the original on thenewstack.io
- Tags
- Coding agents