Sideband Lab
Each computing era optimized its infrastructure around a smaller unit of work. Agents reversed the trend.
For three decades, infrastructure got faster by getting smaller. Monoliths became microservices. Pages became API calls. Functions became invocations. Every generation shaved orders of magnitude off the primitive.
Agents broke the pattern. A single agent task—writing code, researching a topic, executing a multi-step workflow—runs for minutes to hours. Infrastructure now has to hold state, manage context, and stay alive for durations we haven’t seen since the terminal era.
The y-axis shows how long a single infrastructure primitive lasts in each era, on a logarithmic scale. A “primitive” is the atomic unit of work that the infrastructure of that era was optimized around.
Terminal sessions lasted 30 minutes. Microservice hops lasted 10 milliseconds. The entire history of computing infrastructure bent toward making that unit smaller and faster—until agents reversed the curve.
Terminal session duration from Dartmouth Time-Sharing System (DTSS) usage studies. Web browsing from Mozilla Metrics 2010 median session length. Web app response times from Google Analytics industry benchmarks. Mobile API latency from platform median response times circa 2012.
Microservice hop duration from industry P50 latency data. Serverless invocation from Datadog State of Serverless 2020 (Lambda median 800ms). Copilot inference turns from SimilarWeb/Elfsight 2025 ChatGPT session analysis. Agent task duration from METR Time Horizons benchmark, February 2026.