Code output doubled every month for six months, and nobody read ten times more code
Jeff An, CTO of the testing startup Momentic, in Forbes Technology Council: when agents write in parallel, verification moves from reading diffs to proof artifacts, several independent checks, and spending expensive test runs at pre-merge instead of on every push.
Jeff An's Forbes Technology Council piece starts from the flipped ratio: code used to be expensive to write and cheap to review, and agents running in parallel reversed that. Even at a constant bug rate, more output means more bugs reaching production. Momentic moved from mostly IDE coding to Devin, Claude Code and Codex in early 2026, and An says its code output "has at least doubled for each of the last six months" while holding 99.99% uptime. His point is that they didn't get there by asking engineers to read ten times as much code.
Three practices did the work:
- Proof artifacts beside the diff. For real product changes, the agent hands over a short screen recording, screenshots or end-to-end results. A two-minute clip of the agent clicking through the new flow tells a reviewer more than twenty minutes in a large diff. The reviewer starts from evidence and only digs into code when something looks off.
- More than one kind of check. AI review alone misses a lot, and An cites a study where review models did far worse on real pull requests than on synthetic bugs and almost entirely missed performance defects. Queries that fail at scale, services that break only when integrated, and permission bugs tied to one account state don't show up in static reading, so AI review gets paired with AI testing, bug bashes, observability and security scanning.
- Not every check on every push. Agents change code fast enough that the full suite on each commit burns compute and slows feedback. Linear's test suite nearly quadrupled during 2026, and collapsing seven small CI jobs into two saved about 87,000 runner-minutes a month. Cheap deterministic checks stay automatic, agents get skills to run the relevant end-to-end tests while they work, and the heaviest suites wait for pre-merge or pre-deploy.
Why it matters for code review
For teams running agents around the clock, the third point is the one that gets skipped. A continuous loop that pays the full verification bill on every push either stalls or gets its gates quietly removed. An's version keeps the gate by scheduling it: cheap checks constantly, expensive ones where a merge actually happens. The engineer's job shifts from checking every line to designing the system that produces evidence someone can trust. The numbers are self-reported by a company that sells AI testing, so read them as one team's account rather than a benchmark.