Claude Code in Large Codebases: Anthropic insists on the harness, not the model

Anthropic shows Claude Code performance in large codebases depends on a five-block harness, including CLAUDE.md and MCP, managed by an agent manager.

Anthropic is releasing the first installment of a series dedicated to the use of Claude Code in large codebases: multi-million-line monorepos, legacy systems in C, C++, or Java, or constellations of microservices spread across dozens of repositories. The publisher consolidates the patterns observed among its major clients and advocates a thesis: the perceived performance of Claude Code depends more on the ecosystem built around the model, which Anthropic calls a harness, than on the model itself.

The harness is structured around five building blocks. CLAUDE.md files provide the context automatically loaded with each session, one at the root for an overview, others in subdirectories for local conventions. Hooks automate deterministic behaviors (linting, formatting, context updates throughout sessions). Skills offer specialized expertise on demand, avoiding saturation of the general context. Plugins distribute a proven setup to an entire organization. MCP servers connect Claude to internal tools: documentation, ticketing, analytical platforms.

Two capabilities complete the set. LSP integration gives Claude the symbolic precision of an IDE, a critical point for multi-language codebases. Sub-agents allow for separating exploration from editing by isolating contexts.

The article also focuses on the organizational dimension: an agent manager role, a hybrid between PM and engineer, emerges to manage the Claude Code ecosystem internally. The most successful deployments invest in this infrastructure before opening it up to all teams, with a dedicated point person or small team and early dialogue between engineering, information security, and governance.