Simon Willison (Django co-founder, now focused on AI engineering practices) published a subagent engineering patterns guide that breaks down LLM sub-agents into three modes: Serial (sequential exploration), Parallel (concurrent execution), and Specialist (expert roles). The core issue is that context windows are limited, so subagents use independent windows to distribute the load.

In practical application, Parallel is severely underestimated. Dispatching a batch of lightweight models to run exploration concurrently, with the main window waiting for results to return—the experience and efficiency difference is obvious. But the biggest anti-pattern isn't over-decomposition; it's over-trust—results returned by subagents must be cross-validated by the main agent and cannot be directly accepted.

The biggest gain isn't speed; it's context protection. A clean main window allows complex tasks to cross the finish line.

Simon's original link: