Large Language Models (LLMs) face a persistent bottleneck: the context window, which limits how much information they can process at once. Despite rapid intelligence gains, these "working memory" limits often top out around one million tokens, with performance degrading long before that limit is reached. Simon Willison (co-creator of the Django web framework) explores subagents as a critical engineering pattern to solve this. By dispatching fresh copies of themselves with unique prompts, primary agents can tackle complex tasks without saturating their main context window.

The Explore subagent in Claude Code serves as a prime example of this modular approach. When tasked with a new project, the system spawns a subagent specifically to map out file structures and find relevant code blocks. This subagent returns a concise summary to the parent, effectively distilling massive amounts of repository data into a few high-value pieces of information. This hierarchy ensures the main agent remains focused on the core objective rather than getting lost in the technical noise of a large codebase.

Beyond simple exploration, this pattern allows for parallelization and specialization. Developers can run multiple subagents simultaneously—potentially using cheaper, faster models—to update several independent files at once. Specialized subagents can also act as dedicated code reviewers, test runners, or debuggers. While over-engineering these hierarchies is a risk, the primary benefit remains context management, allowing AI to handle larger, more intricate software projects with higher precision and speed.

Large Language Models (LLMs) face a persistent bottleneck: the context window, which limits how much information they can process at once. Despite rapid intelligence gains, these "working memory" limits often top out around one million tokens, with performance degrading long before that limit is reached. Simon Willison (co-creator of the Django web framework) explores subagents as a critical engineering pattern to solve this. By dispatching fresh copies of themselves with unique prompts, primary agents can tackle complex tasks without saturating their main context window.

The Explore subagent in Claude Code serves as a prime example of this modular approach. When tasked with a new project, the system spawns a subagent specifically to map out file structures and find relevant code blocks. This subagent returns a concise summary to the parent, effectively distilling massive amounts of repository data into a few high-value pieces of information. This hierarchy ensures the main agent remains focused on the core objective rather than getting lost in the technical noise of a large codebase.

Beyond simple exploration, this pattern allows for parallelization and specialization. Developers can run multiple subagents simultaneously—potentially using cheaper, faster models—to update several independent files at once. Specialized subagents can also act as dedicated code reviewers, test runners, or debuggers. While over-engineering these hierarchies is a risk, the primary benefit remains context management, allowing AI to handle larger, more intricate software projects with higher precision and speed.