Anthropic outlines enterprise scaling strategies for Claude Code

Anthropic has released a comprehensive guide detailing best practices for deploying Claude Code within large-scale engineering organisations. The publication, titled "How Claude Code works in large codebases: Best practices and where to start," forms part of a new series aimed at addressing the technical and organisational challenges of using the tool in enterprise environments. The guide specifically targets engineering teams managing multi-million-line monorepos, legacy systems, and distributed architectures spanning dozens of repositories.

The article argues that successful deployments rely on a structured "harness" architecture rather than model benchmarks alone. This framework comprises five core extension points: CLAUDE.md files for context, hooks for self-improvement and enforcement, skills for on-demand expertise, plugins for distribution, and Model Context Protocol (MCP) servers for external tool integration. Anthropic emphasises that these components must be layered correctly to optimise performance, with CLAUDE.md files serving as the foundational context layer that loads automatically at the start of every session.

A central technical recommendation is the shift from Retrieval-Augmented Generation (RAG) to "agentic search." Anthropic contends that RAG-based retrieval often fails in large codebases due to embedding pipeline latency and index staleness, which can result in the AI referencing deleted modules or renamed functions. By contrast, agentic search navigates the live local file system, ensuring the tool operates on current code. However, this approach requires significant upfront investment in codebase setup to provide the agent with sufficient starting context to navigate effectively.

The guide also highlights organisational governance as a critical factor in adoption. Anthropic recommends establishing dedicated developer experience teams or appointing an "agent manager" to oversee configuration reviews, plugin distribution, and governance. The firm advises that these configurations should be reviewed every three to six months to ensure they remain effective as underlying models evolve. Without centralised oversight, Anthropic warns that knowledge may remain tribal and adoption could plateau due to fragmented conventions.

While the guidance assumes conventional software engineering environments using Git and standard directory structures, it notes that Claude Code performs well in languages not typically associated with AI coding tools, such as C, C++, C#, Java, and PHP. The publication acknowledges that non-traditional setups, such as game engines or non-Git version control systems, require additional configuration and will be addressed in future installments of the series.