The test-time compute revolution promised better agent decisions—at the cost of latency. But what if we could have both? Sleeper agents represent a paradigm shift: autonomous systems that reason during idle time (non-inference periods), precomputing insights that enable instant responses while maintaining or improving quality. Unlike test-time scaling, which forces users to wait, sleeper agents push reasoning to background compute. A lightweight serve agent answers queries instantly using precomputed context. Result: 1/5 token reduction with 15% accuracy improvements, and 2-3x cost amortization on multi-turn conversations. But this raises new challenges: How do you precompute reasoning without knowing future queries? What representations persist best in memory? How do you prevent hallucination propagation? This talk explores architecture (memory abstraction, precomputation, coordination), challenges (speculative reasoning, eviction, staleness, hallucination), implementation patterns (sleep triggers, refresh mechanics), and production case studies (latency reduced 40%, costs cut 50%). This is frontier infrastructure—barely covered in academic literature, no conference talks to date.

Test-time compute improved agent decisions—at the cost of latency. Sleeper agents invert this: autonomous systems reasoning during idle time, precomputing insights for instant responses. Result: 1/5 token reduction, 15% accuracy gains, 2-3x cost amortization. This talk explores three architectural challenges (speculative reasoning, memory representation, context refresh), production patterns, and case studies of teams reducing latency 40% while cutting costs 50%.