The article introduces the concept of 'subagents' as a pattern in agentic engineering to overcome LLM context window limitations. It explains how coding agents like Claude Code use subagents to delegate specific tasks with fresh context windows, enabling more efficient handling of complex problems without exhausting the primary agent's memory.
Background
Large language models have limited context windows (typically up to 1M tokens) that constrain their working memory, requiring careful management for optimal performance.
- Source
- Simon Willison
- Published
- Mar 17, 2026 at 08:32 PM
- Score
- 6.0 / 10