Jonathon's AI Wiki

Why Coding Is Solved, and What Comes Next — Boris Cherny (AI Ascent 2026)

7 min read

Source: raw/boris-cherny-ai-ascent-2026-transcript.md (YouTube SlGRN8jh2RI, Sequoia Capital / AI Ascent 2026, ~24m36s, published 2026-05-04; Boris Cherny interviewed by Lauren Reader) Type: Fireside / conference talk Speaker: Boris Cherny — creator & Head of Claude Code, Anthropic Event: Sequoia AI Ascent 2026

Boris Cherny’s AI Ascent 2026 fireside, where he argues coding is effectively “solved” (for him, the model writes 100% of his code) and walks through what comes next: phone-first development driving hundreds-to-thousands of agents via loops and routines, cross-disciplinary generalist teams, a Hamilton-Helmer 7 Powers read on which business moats AI erodes, and a printing-press analogy for software’s coming democratization. It’s the same event series as Karpathy’s AI Ascent talk; it overlaps with his Lenny’s Podcast interview and Reflecting on a Year of Claude Code on the “coding solved” + printing-press themes, but adds the SaaS/7-Powers analysis, the phone-first loops setup, and the org thesis.

Key Takeaways

  • “Coding is solved” — for him, literally 100%. The model has written 100% of his code since Oct/Nov 2025; he ships “a few dozen PRs every day,” with a record of 150 PRs in one day. Caveat: not everywhere — big complex codebases and “weird languages the model’s not good at yet”; the usual fix is “wait for the next model.”
  • Claude Code was an accidental bet on a product overhang. Started late 2024 in the Anthropic Labs incubator (which also produced MCP + the desktop app). The thesis: the model could do far more than the state-of-the-art “type-ahead” (Sonnet 3.5 tab-completion) — so let the agent write all the code. It “really didn’t work for the first 6 months” and wasn’t an initial hit; exponential growth started with Opus 4 in May and re-inflects with every model release (4 → 4.5 → 4.6 → 4.7). The team disbanded, then reformed “round two” under Mike Krieger.
  • Why TypeScript + React: the Claude Code codebase is deliberately “on distribution for the model” — when they started, language/framework choice mattered a lot because the model was weaker. (Today the model “can write whatever.“)
  • Phone-first, loop-driven setup. Most of his work now runs from the Claude app on his phone: ~5–10 sessions, a few hundred agents live (a few thousand overnight). His favorite primitive is /loop — Claude uses cron to schedule a repeating job. He runs dozens: babysitting PRs (fix CI, auto-rebase), keeping CI healthy (flaky tests), clustering Twitter feedback every 30 min. “Loops are the future.” Routines are the same thing server-side (close your laptop, it keeps going).
  • Teams trend toward cross-disciplinary generalists. On the Claude Code team, “every single person writes code” — EM, PM, designers, data scientist, finance, user researcher. The next wave isn’t generalist-within-engineering but generalist-across-disciplines (eng + design + data science).
  • The SaaS question, via Hamilton Helmer’s 7 Powers: AI erodes switching costs (port between tools via the model) and process power (Claude 4.7 can “hill-climb anything” — give it a target, it iterates until done). It does not change network effects, scale economies, or cornered resources. Net: expect ~10× more disruptive startups because incumbents face internal change-resistance while a fresh startup builds AI-native from scratch. “Best time to build.”
  • Software is about to democratize like literacy after the printing press. 1400s Europe: ~10% literate → printing press → more literature published in 50 years than the prior 1,000, book cost down ~100×, literacy eventually ~70%. Software will follow the same curve “much faster than 50 years.” Corollary: the best person to write accounting software is a great accountant — “coding is the easy part; knowing the domain is the hard part.”
  • Internal vs external gap: none on the model, large on the org. Anthropic dogfoods the same models everyone gets (mythos to trial, Opus 4.7 for most code). The real lead is organizational: “we use Claude for literally everything,” Claudes talk to each other over Slack, no manually-written code anywhere, all SQL is model-written.
  • Local vs cloud, MCP vs computer use — “it doesn’t matter.” Within a couple years the model decides (it’ll start the agents and build the environments). For knowledge work, Cowork’s answer is the same MCP connectors as Claude.ai (Salesforce, Google Docs/Calendar); computer use is the catch-all for tools without an MCP. “To the model, it’s just tokens.”

Model vs. Product

  • Asked how much of Claude Code’s success is model vs. product: “it’s a mix,” roughly 50/50 both a year ago and six months ago. The YC lesson still binds — “build something people love” — which is why they obsess over the little details.
  • As the model improves, the harness matters less. The current product question is how to evolve the harness — make loops first-class, make it easy to run many agents. In a year, today’s safety scaffolding (prompt-injection defenses, static command verification, permission modes, human-in-the-loop) will matter less “cuz the model will just do the right thing.”
  • Multi-agent is increasingly the model’s job, not the user’s. Parallelization comes down to prompting on the product side, but 4.7 already starts loops on its own (e.g., noticing data changes over time and offering a 30-min report via the Slack MCP). If users have to hand-decide when to parallelize, “it’s a product design problem… I’m not doing a good job.”

Try It

For WEO Marketly / any team reading the roadmap signal:

  1. Adopt loops for the bookkeeping. Stand up a [[claude-ai/scheduled-tasks|/loop]] (or cloud routine) for the chores Boris automates: babysit PRs/CI, cluster inbound feedback on a schedule. This is the highest-leverage pattern in the talk.
  2. Let domain experts build. The “accountant writes the accounting software” point applies directly to a dental-marketing agency — your domain experts (not just engineers) are now the best builders of your internal tools.
  3. Run the 7 Powers test on your own moat. Which of your advantages are switching-cost / process-power (eroding) vs. network-effect / scale / cornered-resource (durable)? Reprioritize accordingly.
  4. Close the “org gap,” not the “model gap.” You already have the same models Anthropic ships; the lead they have is process. Put Claude at the center of workflows (echoing Reflecting on a Year of Claude Code).
  5. Connect tools via MCP first. For knowledge-work automation, wire Salesforce/Docs/Calendar via MCP connectors before reaching for computer use.

Open Questions

  • Net-new vs. his other talks. The coding-solved, 100%-Claude-workflow, and printing-press points retread boris-cherny-lenny-podcast + the boris-cherny hub’s Acquired Unplugged notes; the distinctive AI-Ascent material is the 7 Powers / SaaS-apocalypse analysis, the phone-first loops setup, and the cross-disciplinary-generalist org thesis.
  • “150 PRs in a day” is a self-reported personal stress-test, not a sustained rate — his typical is “a few dozen PRs every day.” ^[ambiguous]
  • No-manually-written-code-anywhere at Anthropic is a strong first-party claim stated in passing; scope (all teams? all repos?) isn’t specified.

Graph View

Backlinks