SecondBrain
Ask the Brain
Index/Sourceupdated Sat May 09 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Boris Cherny on Coding Is Solved (Sequoia AI Ascent)

boris-chernyclaude-codeanthropicloopsagentic-engineeringsaas-apocalypse

Boris Cherny on Coding Is Solved (Sequoia AI Ascent)

Boris Cherny, creator of Claude Code at Anthropic, interviewed by Lauren Reeder at Sequoia AI Ascent 2026. Argues that for the code he writes, coding is solved — he hasn't typed a line in 2026, ships dozens of PRs/day from his phone, ran 150 PRs in a single day as a record. The real story is now about loops, sub-agents, and the printing-press-shaped democratization of software.

Key claims

  • Claude Code origin (late 2024): Anthropic Labs incubator. State of the art at the time was typeahead (Sonnet 3.5 enabled tab-complete by line). Boris's bet: the model could do agentic coding before product-market fit existed. Built it pre-PMF for 6 months. PMF inflected with Opus 4 in May, kept inflecting through 4.5 → 4.6 → 4.7.
  • Coding is solved (for him). Stack is just TypeScript + React because it was on-distribution for the model in late 2024. Today he writes 0% by hand. Caveat: not solved for big complicated codebases, weird languages, etc — "usually the answer is wait for the next model."
  • Workflow now: most work from his phone via the Claude iOS app. 5–10 sessions, each with multiple sub-agents → "a few hundred agents" running. Overnight: a few thousand doing deeper work.
  • Loops are the future. /loop schedules a recurring cron job. He has dozens running: babysitting PRs (auto-rebase, fix CI), keeping CI healthy (fix flaky tests), clustering Twitter feedback every 30 min. Also /batch and the new server-side routines (keep going when laptop closes).
  • Loop emerges spontaneously in 4.7. He'll ask Claude to pull data, and it offers: "I noticed the data is changing — I'll start a loop and report every 30 min via Slack MCP." Product design problem becomes "make loops first-class," not "teach users to spawn them."
  • Anthropic dogfoods aggressively. Same models as everyone (Opus 4.7, some Mythos). Internal Claudes talk to each other over Slack while running in loops. "We have no more manually written code anywhere at the company." All SQL written by models.
  • Future of teams = cross-disciplinary generalists. On the Claude Code team: engineering manager, PM, designers, data scientist, finance person, user researcher — everyone codes. Specialists in something, but coding is now table stakes.
  • No SaaS apocalypse, but a power-shift. Acquired podcast's "seven powers" (Hamilton Helmer):
    • Less important under AI: switching costs (you can port), process power (Claude can hill-climb any process).
    • Still matter: network effects, scale economies, cornered resources.
    • 10× more disruptive startups in next decade — they build AI-native from scratch; incumbents face internal change resistance.
  • Printing press analogy. Pre-1400s, ~10% of Europeans were literate; scribes worked for illiterate kings. After the press: 50 years → more literature than the prior 1000 years; 100× cost drop; eventually ~70% global literacy. Software is at that inflection. The best person to write accounting software is a great accountant, not an engineer — domain knowledge is the hard part now.
  • MCP vs CLI vs computer use is "doesn't matter." "To the model, it's just tokens." MCP for systems with one; computer use for the rest; APIs as fallback.
  • Harness will get less important as models get better. Today's safety scaffolding (prompt injection guards, static command verification, permission modes, human-in-loop) all become less load-bearing as model alignment improves.
  • What he'd build today: Claude Design (will get much better), parallel-agent infrastructure (loop, batch, sub-agents), computer use.

Cross-source resonance

  • Printing press analogy echoes literally in Printing Press (Nate Herk video) (different tool, same metaphor → printingpress.dev was named for it).
  • Loops/cron pattern anticipates the "dozens of agents in production" Praveen describes for enterprises — Boris just pointed it at himself.
  • Harness gets less important is the opposite prediction from Praveen, who says harness design quality is the determinant of how far autonomy can go. Worth tracking — these views are reconcilable (Boris: long-term; Praveen: today) but flag-worthy. See contradictions section in Harness (LLM Agents).
  • "Domain expert > engineer for accounting software" echoes Karpathy's "you can outsource thinking, not understanding" from Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent).

Cross-links