The goal of this thesis is to develop a mathematical framework for optimal, accurate, and affordable complexity statistical learning among networks of autonomous agents. We begin by noting the connection between statistical inference and stochastic programming, and consider extensions of this setup to settings in which a network of agents each observes a local data stream and would like to make decisions that are good with respect to information aggregated across the entire network. There is an open-ended degree of freedom in this problem formulation, however: the selection of the estimator function class which defines the feasible set of the stochastic program. Our central contribution is the design of stochastic optimization tools in reproducing kernel Hilbert spaces that yield optimal, accurate, and affordable complexity statistical learning for a multi-agent network. To obtain this result, we first explore the relative merits and drawbacks of different function class selections. In Part I, we consider multi-agent expected risk minimization this problem setting for the case that each agent seems to learn a common globally optimal generalized linear models (GLMs) by developing a stochastic variant of Arrow-Hurwicz primal-dual method. We establish convergence to the primal-dual optimal pair when either consensus or ``proximity" constraints encode the fact that we want all agents' to agree, or nearby agents to make decisions that are close to one another. Empirically, we observe that these convergence results are substantiated but that convergence may not translate into statistical accuracy. More broadly, optimality within a given estimator function class is not the same as one that makes minimal inference errors. The optimality-accuracy tradeoff of GLMs motivates subsequent efforts to learn more sophisticated estimators based upon learned feature encodings of the data that is fed into the statistical model. The specific tool we turn to in Part II is dictionary learning, where we optimize both over regression weights and an encoding of the data, which yields a non-convex problem. We investigate the use of stochastic methods for online task-driven dictionary learning, and obtain promising performance for the task of a ground robot learning to anticipate control uncertainty based on its past experience. Heartened by this implementation, we then consider extensions of this framework for a multi-agent network to each learn globally optimal task-driven dictionaries based on stochastic primal-dual methods. However, it is here the non-convexity of the optimization problem causes problems: stringent conditions on stochastic errors and the duality gap limit the applicability of the convergence guarantees, and impractically small learning rates are required for convergence in practice. Thus, we seek to learn nonlinear statistical models while preserving convexity, which is possible through kernel methods ( Part III). However, the increased descriptive power of nonparametric estimation comes at the cost of infinite complexity. Thus, we develop a stochastic approximation algorithm in reproducing kernel Hilbert spaces (RKHS) that ameliorates this complexity issue while preserving optimality: we combine the functional generalization of stochastic gradient method (FSGD) with greedily constructed low-dimensional subspace projections based on matching pursuit. We establish that the proposed method yields a controllable trade-off between optimality and memory, and yields highly accurate parsimonious statistical models in practice. % Then, we develop a multi-agent extension of this method by proposing a new node-separable penalty function and applying FSGD together with low-dimensional subspace projections. This extension allows a network of autonomous agents to learn a memory-efficient approximation to the globally optimal regression function based only on their local data stream and message passing with neighbors. In practice, we observe agents are able to stably learn highly accurate and memory-efficient nonlinear statistical models from streaming data. From here, we shift focus to a more challenging class of problems, motivated by the fact that true learning is not just revising predictions based upon data but augmenting behavior over time based on temporal incentives. This goal may be described by Markov Decision Processes (MDPs): at each point, an agent is in some state of the world, takes an action and then receives a reward while randomly transitioning to a new state. The goal of the agent is to select the action sequence to maximize its long-term sum of rewards, but determining how to select this action sequence when both the state and action spaces are infinite has eluded researchers for decades. As a precursor to this feat, we consider the problem of policy evaluation in infinite MDPs, in which we seek to determine the long-term sum of rewards when starting in a given state when actions are chosen according to a fixed distribution called a policy. We reformulate this problem as a RKHS-valued compositional stochastic program and we develop a functional extension of stochastic quasi-gradient algorithm operating in tandem with the greedy subspace projections mentioned above. We prove convergence with probability 1 to the Bellman fixed point restricted to this function class, and we observe a state of the art trade off in memory versus Bellman error for the proposed method on the Mountain Car driving task, which bodes well for incorporating policy evaluation into more sophisticated, provably stable reinforcement learning techniques, and in time, developing optimal collaborative multi-agent learning-based control systems.