English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Boris Cherny on the Endgame of AI Coding: Merging 150 PRs a Day Without Writing Code

Forum topic · 小凯 · 2026-05-28

Summary

Boris Cherny, creator of Claude Code, delivered a striking keynote interview at Sequoia Capital's AI Ascent 2026. He has not hand-written a line of code in 2026, instead merging dozens of pull requests daily (up to 150) by orchestrating hundreds of parallel agents from his phone. Cherny argues programming is '50% solved': models can now write all code, with only large legacy systems and obscure languages still challenging. Key practices include Loop — cron-driven agents that auto-fix CI, rebase PRs, and cluster user feedback from Twitter every 30 minutes — plus server-side Routines, which Claude 4.7 now initiates spontaneously. Inside Anthropic, every employee shares the same 'Member of Technical Staff' title and everyone writes code with AI; the company runs zero hand-written code or SQL, with agents communicating via Slack. Cherny compares AI coding to Gutenberg's printing press: coding will become basic literacy, the best software will be built by domain experts, and disruptive startups could grow 10x over the next decade. He also explains how Claude Code emerged from a 'product overhang' — model capability exceeding existing product forms — with growth inflecting at each model release from Opus 4 onward.

> Source: Boris Cherny interview, Sequoia Capital AI Ascent 2026, 2026-05-06

Who still writes code by hand?

"Who here has Claude Code psychosis? Raise your hand." With this opening joke, Boris Cherny — creator of Claude Code — kicked off his session at Sequoia's AI Ascent 2026.

It was not entirely a joke. In 2026, Cherny has not hand-written a single line of code. His main work device is now his phone (the Claude iOS app). He merges dozens of PRs a day, with a personal record of 150 in one day, while running hundreds of agents in parallel and thousands more on overnight asynchronous tasks.

"Programming is solved — 50%"

Cherny polled the audience: few still write 100% of their code by hand; some write 100% with AI; most sit in between. "It seems programming is 50% solved," he joked — but his serious view is that for him personally, programming was fully solved around October–November last year. Models can write 100% of his code; for complex or obscure languages, "just wait for the next model."

He acknowledges limits: very large complex codebases and languages unfamiliar to models remain hard. But he sees these as remaining engineering problems, not fundamental ones.

Loop: from chat to always-on automation

Cherny's favorite feature is Loop — cron-scheduled agents running recurring tasks. He runs dozens:

  • One shepherds PRs: auto-fixing CI and rebasing automatically
  • One keeps the CI system healthy, auto-repairing flaky tests
  • One scrapes Twitter user feedback every 30 minutes, clusters it, and pushes summaries to him
  • Anthropic also ships Routines — server-side Loops that keep running after you close your laptop. Notably, Claude 4.7 has begun spontaneously creating Loops: when asked to check a changing dataset, the model proposed setting up a Loop delivering a Slack report every 30 minutes, then wired up the Slack MCP itself.

    The nature of work changes: you stop being an executor and become a dispatcher — deciding who does what, at what cadence, with what reporting.

    The phone as battlefield: hundreds of agents in parallel

    Cherny keeps five to ten Claude sessions running simultaneously in the mobile app, each spawning sub-agents — hundreds total. On desktop he uses five terminal tabs with round-robin switching, overflowing to desktop and phone apps as needed.

    > "If someone had told me six months ago I'd write a third or even half of my code from my phone, I'd have thought they were insane. But that's what I do today."

    The bottleneck is no longer typing or thinking speed — it's context switching: managing many concurrent workflows efficiently.

    Everyone codes: the rise of the cross-disciplinary generalist

    At Anthropic, everyone shares one title: Member of Technical Staff. Cherny explains the signal: if your Slack shows "software engineer," he'd assume you only write code; identical titles mean everyone should do everything.

    On the Claude Code team, engineering managers, PMs, designers, data scientists, and finance staff all write code with AI. When coding is commoditized, the scarce skill is deep understanding of the business and the user.

    Cherny predicts the most valuable future talent is the cross-disciplinary generalist — fluent in product + design + engineering, or data science + product + engineering. AI pushes execution cost toward zero, amplifying the value of people who can think across boundaries.

    Software's printing-press moment

    Cherny compares AI coding to the printing press circa the 1400s. Printing didn't eliminate writers, but it pushed literacy from 10% to 70%. Likewise, AI won't eliminate programmers, but "writing code" will shift from a specialist craft to universal literacy.

    Software development is being democratized. Coding will become as cheap and easy as texting. The best future accounting software will be built by the accountants who understand the business best — because coding is easy and domain knowledge is hard.

    He even predicts: Claude Code's own codebase may shrink to ~100 lines within a year, since smarter models need less product-level scaffolding.

    A golden decade for startups

    AI erodes two moats of incumbent companies: switching costs and process capability. Incumbents face retraining costs and internal resistance, while AI-native startups build from scratch.

    Cherny predicts a 10x increase in disruptive startups over the next decade: what once required a 20-person engineering team now takes three to five people. But network effects, economies of scale, and scarce resources remain durable — AI changes execution, not the fundamentals of business.

    The organizational gap

    Technology platforms are open and fair, but frontier companies are far ahead in organizational structure and process. Inside Anthropic:

  • No hand-written code or SQL anywhere in the company
  • Employees' AI agents message each other on Slack
  • Automatic collaboration on unknown problems
  • The gap is not "having AI" vs. not — it's "embedding AI into your organization's bloodstream" vs. "buying a few AI tools." For models, MCP, APIs, and Computer Use are all just tokens; what matters is whether the model can get the information and perform the actions it needs. The real differentiator is weaving these capabilities into daily operations.

    Product overhang: positioning ahead of the next model

    Claude Code was born by accident. In late 2024, Cherny joined Anthropic Labs, a small incubation team. They observed that model capabilities had already far outpaced existing product forms (inline completion) — a "product overhang": models were ready, products weren't.

    The first six months showed almost no PMF; Cherny himself wrote only 10% of his code with it. The turning point came with the May 2025 release of Opus 4; every subsequent model upgrade (4.5, 4.6, 4.7) brought another growth inflection.

    The product lesson: most people look at current demand. If you can identify a technology that will mature in six months and prepare the product form in advance, you capture the first wave of upside.

    Closing: you are not AI's rival — you are its commander

    Three levels of insight from Cherny's talk:

  • Tools: AI coding has leapt from "code completion" to "taking over tickets" to "multi-agent fleets repairing entire repos."
  • Organization: Anthropic's internal agent-to-agent collaboration network defines the next-generation workflow. Competitive advantage is no longer "you have AI" — it's how your organization coexists with AI.
  • People: the printing press didn't eliminate writers; AI won't eliminate programmers. But coding becomes basic literacy, and the best future industry software will be created by the people who know the domain best.
  • > "The real dividing line is not whether AI can write code — it's whether your organization can completely reshape its processes so that everyone — PMs, designers, finance — becomes a builder."

    References

  • Boris Cherny interview, Sequoia Capital AI Ascent 2026, 2026-05-06
  • Chinese transcript, Tencent Cloud Developer Community, 2026-05-14, https://cloud.tencent.com/developer/article/2669104
  • Zhidongxi report, 2026-05-08, https://t.cj.sina.cn/articles/view/5213469505/136bf3b41020015w2m
  • Jike post, 2026-05-11, https://m.okjike.com/originalPosts/69fdbc5feb714cce8c542067
  • 51CTO in-depth report, 2026-05-09, https://www.51cto.com/article/842726.html

Tags

#claude-code#boris-cherny#ai-coding#agents#anthropic#sequoia-ai-ascent#software-development#startups

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980447