Jeff Dean: Agent reliability is a systems problem, not a model problem

Jeff Dean argues the real AI bottleneck has shifted from model capability to system-level orchestration. Agents can now run for days or weeks on complex tasks, but reliability degrades after 10-30 steps. The solution lies in better workflow design, not better models — specifically, giving agents skills and hints that keep

them within their competency distribution. The TPU origin story highlights a key insight: inference latency and energy efficiency will drive the next wave of hardware specialization. Dean notes a 1000x energy gap between computation and data movement, meaning batch processing remains necessary for efficiency but problematic for low-latency inference. This tension

shapes both hardware design and algorithm choices. Dean's "1% rule" for startups is practical: pick problems where current models fail 0-1% of the time, not 20%. If models partially succeed, the capability will likely improve soon. He also emphasizes that clear specifications have become more important with agents, not less —

models need precise direction to avoid inferring wrong objectives. The most provocative insight is about transistors: Dean suggests questioning the assumption that every chip must be identical and error-free. Building systems from unreliable components with redundant pathways could radically change hardware design, echoing how distributed systems handle reliability at scale.