AI coding agents degrade in quality as context windows fill up — not because they run out of tokens, but because their effective 'instruction budget' gets overwhelmed by noise. The key insight is distinguishing between context window size and reliable instruction-following capacity, with the 'smart zone' typically sitting within the first ~100k tokens. Practical strategies include using /clear and /compact commands intentionally, treating plans as handoff artifacts rather than continuing in the same session, and decomposing complex features into agent-sized work items. Each work item should be small enough to complete in one focused session, implemented by a fresh agent starting from a persisted spec. This approach is embodied in the Agentic Coding Toolkit, which structures the workflow around spec creation, decision capture, work item decomposition, and per-item context resets.
