AI Agents flashcards, generated for you.

A system that observes its environment, makes decisions, and takes actions to achieve a goal. Think: a chess player looking at the board, thinking about moves, and playing.

Agents interact with environments over time and learn from consequences of their actions. A single model can't adapt; an agent tries something, sees the result, and adjusts—like learning to ride a bike by falling and improving.

Numerical feedback that tells the agent how good or bad each action was. Positive reward for good outcomes, negative for bad. The agent learns to maximize total reward over time.

1) Observe state (where am I?), 2) Choose action (what should I do?), 3) Receive reward and next state (what happened?). Repeat to learn which actions lead to high rewards.

A strategy or rule that maps states to actions. In other words: given what you observe, here's what you should do. Can be simple (if-then rules) or complex (neural network).

Exploration: trying new actions to discover better strategies. Exploitation: using actions you already know are good. Agents must balance both—explore too much and waste time; exploit too much and miss better solutions.

An estimate of the expected total future reward starting from state s. Intuitively: 'how good is this situation?' A state with high value usually leads to good outcomes if you act optimally.

Q-learning learns Q(s,a) = expected future reward for taking action a in state s. This is more actionable than just knowing state values, because it directly guides which action to pick in each situation.

A mathematical framework where the next state depends only on the current state and action, not on history. This assumption lets us build efficient learning algorithms because the future is conditionally independent of the past.

LLM agents use language models to generate reasoning and actions, then execute steps in a loop (think→act→observe→repeat). They leverage pre-trained knowledge rather than learning from scratch, and can handle complex, open-ended tasks without task-specific reward signals.