AI Agent Latency: Turn Wait Time Into Deep Thinking

When engineers hand off tasks to coding agents like Claude Code, the resulting wait — minutes to nearly an hour — creates a real productivity dilemma. Two dominant responses have emerged: scale up with agent fleets, as in Steve Yegge's Gas Town orchestrator running dozens of instances at once, or protect cognitive health, following a BCG Henderson Institute study that found supervising AI causes 'brain fry' — more major errors and decision fatigue than actually doing the work.

This synthesis argues for a third path: treat agent latency not as dead time but as a resource for deeper thinking, distinct from passive monitoring. Rather than multitasking across agents, engineers should multitask across levels of abstraction — switching to an entirely different project or moving up to architecture and planning while an agent executes. The approach draws on Rich Hickey's 'Hammock Driven Development' and incubation research, arguing that unstructured latency windows are ideal for loading hard problems into background cognition.

The piece also addresses the cost of task-switching itself, via Sophie Leroy's concept of 'attention residue,' and proposes clean handoffs as the fix: delegation that ends in a written spec is a completion, not an interruption — which aligns with Spec Driven Development tools like GitHub's Spec Kit. For teams increasingly built around AI coding agents, deliberately structuring the wait is becoming as important as writing the prompt itself.