Anthropic's Boris Cherny Ships 20-30 PRs Daily Using Parallel Claude Code Instances
Key Takeaways
- ▸Boris Cherny ships 20-30 PRs daily using five parallel Claude Code instances with a plan-first, one-shot implementation workflow
- ▸Simple glob and grep operations outperformed RAG and vector databases for agentic code search in production use
- ▸Claude Cowork was built in ~10 days and is growing faster than Claude Code, targeting non-engineer users with enhanced safety features
Summary
Boris Cherny, creator and Head of Claude Code at Anthropic, revealed in a detailed podcast interview how he uses AI-powered coding tools to dramatically scale his development output. Cherny ships 20-30 pull requests per day by running five parallel Claude instances across separate terminal tabs, using a plan-first workflow where the AI one-shots implementations after iterating on the plan. The approach represents a fundamental shift in how software engineering work is performed at AI-native companies.
Claude Code's technical architecture favors simplicity over complexity. The team discovered that basic glob and grep operations, driven by the model, outperformed sophisticated approaches like RAG and vector databases for code search. This decision was inspired by observing how engineers at Instagram searched code when IDE features failed. The tool evolved from an internal side project into a core development platform at Anthropic, where the company's flat structure—everyone holds the title "Member of Technical Staff"—encourages cross-functional work.
The team recently built Claude Cowork in approximately 10 days, targeting non-engineers who were already using Claude Code. Cowork is growing faster than Claude Code did at launch, addressing latent demand from data scientists, finance, and sales teams. Most engineering effort went into safety features: classifiers, sandboxed VMs, OS-level protections against file deletion, and rethinking permissions for non-technical users. Cherny emphasized lessons from Meta showing that code quality has measurable double-digit percentage impacts on productivity—a principle that applies equally to AI-generated code.
- Anthropic uses a flat title structure where all employees are "Member of Technical Staff," encouraging cross-functional collaboration
- Code quality has measurable double-digit impacts on productivity for both human and AI-generated code, based on Meta research
Editorial Opinion
The success of Claude Code's deliberately simple architecture—choosing glob and grep over vector databases—challenges the industry's reflexive reach for complex RAG solutions. Cherny's workflow of managing five parallel AI coding agents represents an early glimpse of how senior engineers may evolve into orchestrators rather than implementers. The rapid 10-day development of Claude Cowork and its faster-than-expected adoption suggests the真正 revolution isn't replacing programmers, but democratizing software creation across business functions—a shift with profound implications for how companies structure technical work.
