Anthropic's Claude Code Chief Says He Hasn't Written Code by Hand in 8 Months, Relies Entirely on AI

Boris Cherny, head of Anthropic's Claude Code, says he hasn't written code by hand in eight months, instead managing thousands of AI agents that now write the tool themselves. His role has shifted to "loop engineering," designing automated systems where AI writes, reviews, and improves code. Cherny compares this to the printing press, arguing it democratizes software creation but also creates new bottlenecks and organizational challenges.

AI-assisted software development

• ▪Boris Cherny's role has evolved from prompting AI to a practice called "loop engineering," where he writes loops that orchestrate the agents' work.
• ▪The Claude Code tool is now entirely written by the AI model Claude Code itself.
• ▪Boris Cherny, the head of Claude Code at Anthropic, stated he has not written a line of code by hand in approximately eight months.
• ▪Instead of writing code, Boris Cherny's work now involves managing hundreds to tens of thousands of autonomous AI agents.
• ▪Boris Cherny compared the impact of AI coding assistants on software creation to the invention of Gutenberg's printing press on literacy.

Dynamic workflow architecture

• ▪Loop engineering is an orchestration pattern that combines scheduled execution, isolated workspaces, verifier agents, and persistent memory.
• ▪Anthropic uses a system where AI agents called "sub-agents" are prompted by other Claude instances, making human prompting an exception.
• ▪A dynamic workflow moves the plan for a complex task into code, so the AI only needs to handle the current step, avoiding context window limitations.

Context window limitations

• ▪Automating one part of a process, like code writing, can create a new bottleneck in a subsequent stage, such as code review.
• ▪Anthropic identified three common failure modes when a single AI context window handles a large task: agentic laziness, self-preferential bias, and goal drift.
• ▪To solve the code review bottleneck, Anthropic uses a team of Claude instances with distinct personas to collaborate on reviewing pull requests.

Workflow composition patterns

• ▪Anthropic has identified six common workflow composition patterns, including "fan-out-and-synthesize" and "adversarial verification."
• ▪Developer Jared Sumner reportedly rewrote the Bun JavaScript runtime from Zig to Rust in six days using dynamic workflows, a task estimated to take a year for a human team.
• ▪The "adversarial verification" pattern uses one agent to perform a task and a separate agent to try to disprove or find flaws in the result.

Non-technical workflow applications

• ▪Dynamic workflows can be applied to non-technical problems, such as analyzing a business plan from multiple hostile perspectives simultaneously.
• ▪Anthropic encourages new engineers to engage in peer programming with other humans to foster team connection, as they might otherwise interact mostly with Claude.
• ▪At Fortune's Brainstorm Tech conference, Boris Cherny said Anthropic seeks to hire generalists, people with low ego, and empiricists who learn from data.
• ▪A Harvard Business Review study found that overusing AI tools can lead to mental fatigue called "AI brain fry," causing slower decision-making and more errors.