Anthropic’s Boris Cherny on Claude Code, Agents, and the Next Software Shift

May 4, 2026 · coding · Source ↗

Published 2026-05-04 - Runtime about 25 min - Watch on YouTube

TLDR

Boris Cherny says Claude Code moved from typeahead to agents, and his own workflow now ships dozens of PRs a day from his phone.
He expects software to follow the printing press path: cheaper coding, more literate builders, and far more startups that can outcompete incumbents.

Key Takeaways

Claude Code began inside Anthropic Labs in late 2024, alongside MCP and the desktop app, as a product bet on the next model.
The team built against a 6-month pre-PMF gap because the model could already go beyond typeahead and write whole codebases.
Cherny says Claude Code’s codebase is simple TypeScript and React because those were more on-distribution for earlier models.
He says the model wrote 100% of his code by October or November last year, and he now often uses his phone as the main workspace.
His setup includes 5 to 10 active sessions, a few hundred agents, and a few thousand deeper overnight agents.
His favorite automation is /loop: a cron-like repeat job used for PR babysitting, CI repair, flaky tests, and Twitter feedback clustering.
Anthropic also launched routines, which works like loops on the server so work continues even if the laptop closes.
He predicts more cross-disciplinary generalists, where designers, PMs, data scientists, and finance people all code.
On the Claude Code team, every function he named, including engineering manager and user researcher, writes code.
He thinks switching costs and process power matter less as models improve, while network effects, scale economies, and cornered resources remain important.
He argues AI-native startups can move faster because incumbents must retrain people and change business processes before they can respond.

Notes

Cherny joined Anthropic Labs in late 2024; the team built Claude Code, MCP, and the desktop app, then disbanded.
Mike Krieger now leads the rebuilt Anthropic Labs team for round two.
He says Sonnet 3.5 made typeahead meaningful, but the next step was letting an agent write all the code.
Claude Code’s exponential growth started with Opus 4 in May, then accelerated with 4.5, 4.6, and 4.7.
For him, coding is effectively solved, but not everywhere; large codebases and weird languages still lag.
He says he wrote no code in 2026 and once did about 150 PRs in a single day.
His team uses Claude for everything, and Claude instances talk over Slack to coordinate with other Claude loops.
Anthropic writes all SQL with models and has no manually written code anywhere in the company, per Cherny.
He thinks Claude’s harness will matter less as models improve, while safety checks and permission modes become less central.
He expects the model itself to become better at parallelization, loops, and delegation rather than users orchestrating every sub-agent.
For knowledge work, he points to Salesforce, Google Docs, and Google Calendar as MCP targets; for missing connectors, computer use is the catchall.
His strongest analogy is the printing press: before it, about 10% of Europeans were literate, and afterward book costs fell about 100x and literacy rose to around 70%.