Boris Cherny

Boris Cherny is the creator and head of Claude Code at Anthropic and the author of Programming TypeScript. He builds coding agents that can understand repositories, operate existing development tools, and check their work against tests and other concrete feedback.

Before Anthropic, Cherny worked in finance and at Meta, where he was a technical lead for Facebook Groups before joining Instagram. Working remotely from Japan sharpened his preference for hands-on programming, delegation, and developing other engineering leaders, a philosophy he explored in his essay on remote engineering leadership.

His earlier work focused on making software more precise through programming languages and types. His 2019 O’Reilly book distilled that approach for TypeScript developers, while his open-source projects json-schema-to-typescript and flow-to-typescript translate JSON Schema definitions into TypeScript declarations and migrate Flow-annotated code into TypeScript, respectively.

At Anthropic, Cherny developed Claude Code around a deliberately flexible premise: coding agents should fit the developer’s environment while model capabilities and programming interfaces continue to evolve. The terminal provides direct access to the model, with editors, GitHub workflows, and programmable integrations supporting different ways of working.

• Models as composable Unix utilities. A coding agent can accept piped logs or other inputs, produce structured output, and pass results to ordinary tools such as jq. Shell commands and Model Context Protocol integrations extend its abilities without requiring a custom plugin for every task.

• Verification-driven coding agents. Cherny recommends having an agent explore a repository, gather context, and propose a plan before editing. Tests, browser screenshots, integration results, or simulator output then give it observable targets against which to improve successive attempts.

• Repository context as working infrastructure. Git history, repository instructions, reusable Markdown commands, and existing command-line tools give agents the concrete context needed to reason effectively about unfamiliar codebases.

• Parallel agents with lightweight coordination. Separate checkouts, Git worktrees, and GitHub Actions let multiple agents pursue independent tasks; shared Markdown files can coordinate them when necessary. His public account of his Claude Code setup emphasizes flexible workflows and concurrent sessions.

Cherny’s AI Engineer World’s Fair talk places these choices within programming’s evolution from text editors and autocomplete toward natural-language systems capable of planning, using tools, and verifying their own output.