Code Gen Isn’t the Constraint Anymore. CI Is.

TLDR

• Anthropic says their engineers ship ~8x more code per quarter than before. Claude writes ~80% of that code (at Anthropic, not everywhere).
• Tests ~10x. CI jobs ~25x in six months.
• They tried three patches. Those lasted 70 days, then 29, then under a day. Then they rebuilt.
• Writing code stopped being the scarce step. CI did not grow for free.
• If you add parallel agents, expect more PRs and more CI. Budget for that early.

We’ve spent two years saying AI makes engineers faster. Fine. The part I care about is which queue starts screaming next.

Writing got cheap. Review got faster. Then CI decides if you ship today or stare at a red pipeline.

What Anthropic said

On 14 Sep 2026 they published an eng post: Agentic coding is straining CI.

Their numbers (their company only):

• ~8x as much code shipped per quarter vs their 2021-2025 average
• Claude authors ~80% of that shipped code, and also helps review/approve PRs
• Tests across the codebase ~10x, with only a small increase in engineers
• CI jobs ~25x over six months
• Three patches before a redesign lasted 70 days, then 29, then under 1 day
• The redesign took about 3 weeks and one engineer (they say a similar change used to take about a quarter)

Their line that stuck with me: writing code is no longer the constraint. Once PR review speeds up, CI feels the pressure.

That matches what I see locally too. People talk about vibe coding. The failing part is rarely the first draft. It is the queue after: review, tests, CI, flaky retries.

Also: when their test-selection listeners lag, the risk is stale selection data. More flaky or failing tests get run. They are not saying “we shipped untested to prod.” Keep that straight.

Claude, in their account, likes smaller PRs. Agents keep work moving overnight and on weekends. Humans still approve, so it still comes in bursts. Net: more merges, more jobs.

Around the same week they also talked about Claude Code Projects with parallel threads (blog). I’m not writing a product review. I only care about the load: more parallel agents means more PRs and more CI. If CI was already strained, that is a multiplier.

What I’d budget for

If your team is celebrating “AI writes most of our PRs,” ask one boring question in the same meeting: what is the plan for 10x-25x CI jobs?

I wouldn’t wait for the weekly patch cycle to become the product.

• Treat CI and test selection as capacity, not a weekend cleanup
• Prefer a design that survives big load over patches that buy a few weeks
• Expect smaller PRs and more overnight work if agents are in the loop
• Watch for flaky amplification before you celebrate merge rate
• Don’t buy parallel agent workflows without buying parallel CI

Don’t paste Anthropic’s 80% onto your company. Stacks differ. Review culture differs. The useful part is the pattern: generation scaled, half-measures died faster each time, architecture had to change.

I wrote earlier about review triage signals going weird under AI-shaped PRs (post-028). That was about review. This is about CI after review speeds up. Related. Not the same claim.

Agentic coding did not invent downstream bottlenecks. It just moved them to the front of the room.

Most “AI coding productivity” work is infrastructure work. It just has a nicer name.