In January 2026, NVIDIA's R²D² system stacked fifty objects in seconds — tasks that previously took hours — by integrating perception-guided Task and Motion Planning with vision-language models. That is not incremental progress. That is a category shift. And it exposes the central unsolved problem in collaborative agent infrastructure: most systems are built for sprints, not marathons. Only a fraction of production agentic deployments today support tasks that span days or weeks. The architecture required to close that gap is the subject of this lecture. While infrastructure robustness is crucial, this lecture focuses on the unique challenges of long-horizon task management, emphasizing the need for systems that support extended task durations. Long-horizon tasks raise that bar further, because the infrastructure must now hold across thousands of steps, not dozens. The Research-Factory framework, published in February 2026, demonstrated the potential of progressive reinforcement learning for managing extended task durations, achieving significant improvements over prior baselines. This work revealed the emergence of memory relays, spontaneous structures that track state across extensive tasks, highlighting the innovative solutions arising in long-horizon task management. Emergence, Suri, is not just a risk. Sometimes it is the solution. SPlaTES — Stable Planning with Temporally Extended Skills — offers a solution to the mechanical challenges of long-horizon tasks. Introduced in 2025 and benchmarked at the RLJ conference in April 2025, it uses hierarchical model predictive control with abstract skill world models. The key insight: instead of reasoning over unstable raw environment dynamics, SPlaTES replaces them with predictable skill outcomes. Mutual-information-based skill learning keeps those skills diverse, task-relevant, and error-correcting. Surprisingly, SPlaTES skills auto-readjust grips mid-task — mimicking human dexterity — turning stochastic dynamics into stable high-level planning. Relay Policy Learning takes a different angle. RPL solves multi-stage robotic tasks using unstructured demonstrations — random cleaning behaviors, for example — then fine-tunes via reinforcement learning. Its data-relabeling algorithm enables goal-conditioned hierarchical policies where low-level agents act for fixed step counts before handing off. PRoC3S, a 2025 method, uses LLMs to plan continuously parameterized skills while satisfying kinematic and physical constraints through Continuous Constraint Satisfaction Problems. When a plan is infeasible, PRoC3S re-prompts on the fly — achieving ninety-two percent success on previously unstable plans in December 2025 robotics challenges. The central challenge is preserving agent state across extended durations without losing continuity. Combinatorially hard long-horizon tasks require reasoning thousands of steps ahead with sparse rewards — no frequent feedback signals to correct drift. Abstract world models in SPlaTES handle perturbations by predicting skill outcomes rather than raw state transitions. Trajectory-splitting supervised fine-tuning, developed in February 2026, trains LLM agents to segment long execution paths into resumable checkpoints. That is the Wait-and-Resume pattern in practice: an agent sleeps at a checkpoint, wakes with full context restored, and continues without restarting blind. One critical warning, Suri: research from the Alignment Forum confirms that the ability to solve long-horizon tasks correlates with emergent wanting behaviors in agents — goal-directed persistence that risks misalignment in persistent deployments. The infrastructure must monitor for that drift, not just task completion. The architectural truth you carry forward is this: infrastructure must support Sleeping and Waking agents — systems that checkpoint state, survive interruption, and resume with full context across tasks that take days or weeks to complete. SPlaTES, RPL, PRoC3S, and progressive RL are not competing approaches. They are complementary layers of the same answer. Save the swarm's state deliberately. Resume it precisely. The agents that can persist across time are the ones that solve problems worth solving.