Lenny's Podcast

Claude Code head Boris Cherny: Why he ships 30 PRs a day

Through hundred-percent AI-written code and parallel running agents on autopilot; 'clodify everything,' unlimited tokens, and latent demand make the builder.

Boris ChernyguestLenny Rachitskyhost

Feb 19, 20261h 27mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Claude Code’s head explains agentic coding, builders, and safety ahead

Boris Cherny says he no longer hand-edits code: Claude Code writes 100% of it, enabling him to ship 10–30 PRs per day while still performing human oversight and using Claude for automated code review.
He frames coding as “largely solved” for many real-world stacks and argues the next frontier is agentic work: AI that uses tools, acts in systems, and increasingly suggests what to build by mining feedback, bug reports, and telemetry.
The conversation covers how Claude Code emerged from lightweight prototyping, why the terminal form factor worked early, and how “latent demand” (users misusing a tool in revealing ways) shaped both Claude Code and Cowork.
Cherny also emphasizes safety as Anthropic’s core mission, describing a layered approach (interpretability, evals, and real-world behavior) and advocating for a “race to the top” via open-sourcing safety infrastructure like sandboxes.

IDEAS WORTH REMEMBERING

5 ideas

AI coding is shifting engineers from typing code to directing agents.

Cherny reports Claude Code writes all his code and he hasn’t hand-edited since November, but he still reviews outputs and relies on checkpoints like AI + human review—suggesting the core skill becomes specifying goals, validating results, and managing throughput.

The next leap is AI proposing work, not just executing it.

Claude increasingly generates fix/feature ideas by scanning Slack feedback channels, bug reports, and telemetry, then offering PRs—moving from “assistant” toward “coworker” that helps decide priorities.

Speed comes from under-resourcing and forcing “clodification.”

He argues that small teams (or even one engineer) move faster because they’re compelled to automate via Claude rather than coordinate; the constraint drives creative delegation to agents instead of adding headcount.

Don’t optimize token spend too early—maximize experimentation first.

Cherny advises giving engineers “as many tokens as possible” initially because the cost is small relative to salary at exploratory scale; optimization (cheaper models, tighter loops) should come after an idea proves valuable.

Build products around what users and the model are already trying to do (latent demand).

Claude Code and Cowork both emerged from observing “misuse”—developers using a terminal agent for non-coding tasks (tomato growing, corrupted photo recovery, MRI analysis). The signal: if people jump through hoops, a dedicated product will likely land.

WORDS WORTH SAVING

5 quotes

A hundred percent of my code is written by Claude Code. I have not edited a single line by hand since November.

— Boris Cherny

Productivity per engineer has increased two hundred percent.

— Boris Cherny

Coding is largely solved.

— Boris Cherny

By the end of the year... the title software engineer is gonna start to go away. It’s just gonna be replaced by builder.

— Boris Cherny

The product is the model. We wanna expose it... give it the minimal set of tools, so it can decide which tools to run.

— Boris Cherny

100% AI-written code and multi-agent workflowsProductivity gains and new engineering bottlenecksFrom coding to tool use to computer use (agents)Cowork: non-coding agent automation and desktop/Chrome controlLatent demand as a product discovery methodBuilding for the next model (six months out) and the “bitter lesson”Safety stack: mechanistic interpretability, evals, in-the-wild monitoring

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.