Anthropic's daily merge volume rises to about 30,000: relying on verifiable evidence rather than reading diffs line by line
In a talk at aiDotEngineer, Addy Osmani said Anthropic's daily merge volume rose from about 2,500 in May to about 30,000; the approach is that Claude performs the first round of review, but must attach evidence such as end-to-end, browser, lint, test, screenshot, or video evidence, and humans review the evidence package rather than reading through every diff line by line. The process also includes using hooks to watch edits and diffs, writing team conventions into CLAUDE.md or agents.md, and holding that Claude Opus 5.5 is already sufficient, with the next wave of gains lying in verification practices. CircleCI noted that about 30% of merge attempts still fail, and SonarSource suggested independent verification and guardrail-style constraints at each stage of agent, CI, and maintenance.
Why read this
AI score 64
A share discloses process details of how Anthropic's daily merge volume grew from about 2,500 to about 30,000, showing how verification evidence replaces line-by-line diff review.
Topics
Body · AI translation
Thirty thousand merges a day: Anthropic relies on "verifiable evidence" rather than more humans reading diffs
From a talk by @kapmani and @addyosmani at @aiDotEngineer: when coding agents make "writing code" no longer the bottleneck, what actually holds things up is whether there is enough evidence to dare to merge. Addy says Anthropic's daily merge volume rose from about 2,500 in May to roughly 30,000: Claude does the first round of review, but it must come with evidence such as end-to-end, browser, lint, test, screenshot/video artifacts; humans review the evidence package, not every line of every diff. Manish's closing summary: writing fast doesn't mean shipping fast; autonomy is earned. https://t.co/R62YHtJYDy
What was discussed on stage? A show of hands: almost everyone has had an agent open a PR, but very few actually "read every line by hand before merging." CircleCI's side note: CI throughput has surged, yet about 30% of merge attempts still fail. The framework is the software factory / software production line: in a lit factory, humans still apply judgment and taste across plan–build–check–release; a dark factory is closer to a fully automated line, where humans mainly set intent and handle big exceptions. Most teams sit in the middle of the spectrum.
How does Anthropic handle PRs at this scale? The process: Claude does the first round of review → it tries the product as much like a real user as possible (browser, tools, lint, various checks) → when gaps are found, hooks watch edits and diffs and the PR is fixed as soon as possible → team conventions are written into CLAUDE.md / agents.md and the like → humans check the evidence the agent hands over (including UI screen recordings and screenshots). Addy stresses: just saying "verify" is too vague; you need to get down to concrete test methods such as end-to-end, property, and visual verification before automated merging is even on the table. The Claude Opus 5.5 model is already good enough; the next wave of gains lies more in verification practices and sharing them.
@SonarSource's side: multi-layered verification and guardrails The State of Code direction: most people feel AI gets features shipped faster, PRs are bigger and more complex, and a considerable share of code still reaches production without sufficient review; this is especially glaring in regulated settings. The recommendation is zero trust, with independent verification at every ring (the agent ring, the CI ring, the maintenance ring): the deterministic layer catches injection/overflow and the like; agent review catches logic and intent; the supply-chain layer looks at CVEs and malicious packages. Enterprises also need to manage agent identity, telemetry, permissions, and sandboxing; they prefer to talk about guide rails: keeping agents on approved paths rather than blocking after the fact.
Quick-fire Q&A highlights By the end of 2027, most teams: leaning toward "lit with a few dark corners," unlikely to be all dark. PRs written by agents should be labeled (short-term audits, extra checklists); responsibility still lies with the initiator: if you kicked off Claude and it ran, you're still the one who fixes it when something goes wrong. What juniors/leads should practice this year: figure out what "verification" actually means in terms of tests for your own product. When is it OK to merge: when you have sufficient confidence, or have reduced risk to an acceptable level.