Explainer: エージェント（Agent） splits into reactive, deliberative, hybrid

The explainer frames agents as a concept that has been a central research subject in computer science since the 1990s, with applications spanning robots, game AI, and simulations. Its three-paradigm breakdown traces a progression in how much internal state a system carries: reactive agents decide with no memory, deliberative agents maintain a world model and plan, and hybrid agents switch between the two depending on urgency.

The piece also maps decision-making mechanisms to applications in a table, pairing rule-based methods with game NPCs, optimization with robot control, game-tree search with pre-AlphaGo game AI, machine learning with AlphaGo and autonomous driving, and reinforcement learning with game AI and robots. It then walks through AlphaGo's agent structure as a concrete case, describing neural-network value evaluation, Monte Carlo tree search for candidate moves, and a policy network before the selected move is executed.

Modern examples extend the same loop to large language model agents such as AutoGPT, which take a user instruction and external information as input, decompose it into intermediate tasks, select tools, and plan an execution order before generating text and calling tools. The explainer closes by listing open challenges — transferability beyond a single game, sample efficiency compared with human learning, explainability of chosen actions, and convergence guarantees in multi-agent learning — without offering a timeline for resolving them.