Agentic Coding Is Anything But (matcohle.me)

The author presents a practical framework for "agentic coding" and argues that agents multiply, not replace, human cognitive effort. Using a simple unit model (T05, T15, T40) they show an agent reduces the cognitive load of tasks by an "assist ratio" but only up to an "agent ceiling" — tasks above that ceiling become slower because you spend time fixing and instructing the agent. Crucially, value is generated only when humans invest cognitive capacity, so an agent’s contribution is multiplicative: zero human input → zero value. Tasks that can’t be cleanly decomposed or require holistic reasoning don’t benefit and may get worse when "dissolved" into smaller pieces due to planning and coordination overhead. A concrete example: a T40 task must be split into many smaller T05 pieces for an agent with a ceiling of 10; with an assist ratio of 2 you can finish faster, but dissolve overhead reduces the net gain. The author’s lived estimate is roughly 2× productivity on average, not the dramatic 10× claimed elsewhere, and the real wins come from offloading low-cognitive chores (formatting, trivial fixes) so engineers can focus. Implications: agents push organizations to give engineers uninterrupted focus, favor teams whose work is mostly small, well-scoped tasks, and will improve as the agent ceiling rises — but human cognitive capacity remains the ultimate bottleneck.