The promise of AI-augmented knowledge work was straightforward: automate the tedious parts, reclaim your time, go home earlier. A senior engineering leader at Bold.org has published a detailed personal case study — backed by seven years of GitHub contribution data — that documents the opposite outcome. His output in 2026, while holding a VP-level role that should produce less code than ever, has already surpassed every full year since 2020. He is not going home earlier. He is working more than he ever has. The mechanism is precise and worth understanding. As a VP of Technology, his job shifted years ago from writing features to architecture, planning, and leadership. His GitHub contributions dropped accordingly — from 2,711 in 2020 as a full-stack engineer to 1,189 in 2023 as a principal engineer. Then AI coding agents arrived. By October 2026, he had logged 2,102 contributions, with August 2026 hitting 427 — the single busiest month in his seven-year history. His median weekday contributions jumped from 3-4 (2022-2025) to 7 in 2026. Days with 24+ contributions went from exactly one per year to sixteen. The critical insight is not that the tools are fast. It is that cheap implementation removes the natural friction that previously limited scope. When coding a feature costs hours of focused work, you triage ruthlessly. When an agent can do it in minutes, every marginal idea becomes worth pursuing. Bugs that would have been tolerated get fixed. Patterns that would have waited get refactored. Experiments that would never have started get launched. The workload did not shrink — it expanded to fill the new capacity, and then some. The cognitive load profile has changed shape entirely. Instead of holding one or two problems in deep focus and completing them sequentially, the author now manages several agents in parallel — each requiring intermittent attention for approvals, context switches, code review, and decision-making. The mental cost is not in any single task but in the constant switching between them. Two serious tasks a week was tiring in a familiar, recoverable way. Ten shallow-but-real tasks across multiple agents is tiring in a way that accumulates differently. The data reveals a boom-bust cycle that has repeated three times in 2026. A new model capability or workflow breakthrough triggers a burst of ambition — January-February (277 and 233 contributions), April (187), and August-September (427 and 301). Each burst ends with cognitive exhaustion: the author stops wanting to start agents entirely and retreats to baseline work. But even the valleys are elevated — March 2026 at 203 contributions still exceeded his principal engineer monthly average of 114. The low season is no longer low. This is a single data point, not a study. But the structural pattern it describes — AI reducing per-task cost while increasing total task volume, shifting bottlenecks from execution to attention management — echoes across industries adopting generative AI. The historical precedent is well-documented: the washing machine did not give households more leisure, it raised cleanliness standards. Email did not reduce communication overhead, it multiplied the number of messages requiring response. The question is whether this cycle is a transition cost that resolves as workflows mature, or a permanent feature of tools that eliminate execution friction without expanding human cognitive bandwidth. The author is honest about not having solved this. He likes the expanded capability. He has not figured out how to cap the expanded workload. The tools delivered exactly what was promised — the ability to do more. The part nobody modeled was that 'able to do more' converts almost instantly into 'expected to do more,' including by yourself.
