Head of Claude Code: What happens after coding is solved | Boris Cherny

What It Covers

Boris Cherny, Head of Claude Code at Anthropic, details how Claude Code grew from a two-liked internal hack to generating 4% of all GitHub commits within one year. He covers the shift from AI-assisted coding to fully AI-generated code, what comes after coding is solved, and how agentic AI expands beyond engineering into broader knowledge work.

Key Questions Answered

• •AI Coding Saturation Point: Claude Code now authors 4% of all public GitHub commits, with projections reaching 20% by year-end. Private repository usage is estimated even higher. Anthropic's engineering team has seen a 200% productivity increase in pull requests per engineer since adoption, while the team itself only grew approximately 4x — meaning output scaled far faster than headcount.
• •Token Generosity as Innovation Strategy: Give engineers unlimited token access before optimizing costs. Experimentation at the individual level remains cheap relative to salaries. Only optimize model choice — switching from Opus to Sonnet or Haiku — after a use case proves viable at scale. Restricting tokens early kills the exploratory behavior that produces breakthrough product ideas and workflows.
• •Build for the Model Six Months Ahead: Products built for today's model become obsolete quickly. Scaffolding and workflow constraints may improve performance 20%, but those gains typically disappear with the next model release. Instead, design for anticipated model capabilities — longer autonomous run times, better tool use — accepting weaker product-market fit early in exchange for strong positioning when the next model ships.
• •Latent Demand as Product Signal: When users abuse a product to accomplish tasks it was never designed for — Claude Code users analyzing MRIs, recovering corrupted photos, growing tomatoes — that behavior signals where to build next. Cowork, Anthropic's general agent product, was built in ten days after observing non-engineers jumping through terminal setup hurdles just to access Claude Code for non-coding tasks.
• •Optimal Claude Code Usage Protocol: Use Opus 4.6 with maximum effort enabled — cheaper models often consume more total tokens due to lower accuracy and more correction loops. Enable plan mode (Shift+Tab twice in terminal) for roughly 80% of tasks, letting the model outline steps before executing. Once the plan looks correct, enable auto-accept edits; Opus 4.6 one-shots most tasks successfully from a solid plan.
• •The Generalist Advantage in an AI-Native Workplace: Engineers who cross disciplines — combining product thinking, design sense, business awareness, or user research skills with technical ability — will capture disproportionate value. On the Claude Code team, every role including product management, design, finance, and data science involves coding. The title "software engineer" is expected to fade, replaced by "builder" as role boundaries collapse further by end of 2025.