The second brownfield trick is to use parallel agents: running multiple isolated Claude Code workstreams on different aspects of your codebase.

When you're working on a complex system, you often have parallel workstreams. Maybe you need to refactor a core service, update an API layer, and fix some business logic all at the same time.

Instead of context-switching between these tasks in a single Claude conversation, open separate sessions, subagents, or worktrees - one for each major task. Give each worker a focused instruction file that specifies its mission. One becomes your service refactoring specialist, another handles the API layer, and a third focuses on business logic fixes.

The key insight is that each Claude instance maintains its own context and conversation history. They won't interfere with each other. You can have one doing exploratory analysis while another is writing tests while a third is refactoring code. They all work in parallel.

This is particularly powerful for large brownfield projects where different parts of the system have different contexts and concerns. Instead of cramming everything into one conversation and watching your context window fill up, you distribute the cognitive load across multiple specialized agents.

Just be mindful of potential merge conflicts if they're working on related files. But for the most part, this parallelization dramatically speeds up your work on large, complex codebases.