Karpathy’s Autoresearch gives every experiment exactly 5 minutes of wall-clock time. The question shifts from “what model is best after N steps?” to “what configuration gives the best result within this exact time budget?” That reframing changes the optimization surface entirely — the agent must optimize for improvement per unit time, not abstract model quality.
Time bounds are a resource constraint that shapes behavior, just as The context window is the fundamental constraint — everything else follows shapes how agents plan and decompose work. In both cases, the constraint is not a limitation to work around but a design force that produces better outcomes: context limits force decomposition into small stories, time limits force optimization for real-world throughput. This principle connects to A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one — the time budget is part of the harness, and it does more to ensure practical results than a smarter agent running unbounded would. It also strengthens Autonomous coding loops need small stories and fast feedback to work by adding a concrete mechanism: fixed time = fast feedback guaranteed.