An agent is an LLM-driven system that operates in a loop, where the LLM decides what action to take next based on the current state, and the loop continues until a terminal condition is met.

• The three things that distinguish an agent from a plain LLM call or a RAG pipeline:

LLM → RAG → Agent. A plain LLM is stateless one-shot generation. RAG adds retrieval to ground the generation. An agent adds a decision loop and tools — it can choose what to retrieve, what tool to call, and when to stop.

• 2. The anatomy of a single agent

3. ReAct: The canonical loop

• 4. Tool calling — the action interface
• 5. State and memory
• 6. Other single-agent patterns
• 7. Failure modes (canonical list)
• 8. Production considerations (the FDE angle)
• 9. When single-agent vs multi-agent

An agent is an LLM operating in a loop, where the LLM decides what tool to call next based on the current state, until a terminal condition is met. The four components are the model, tools, memory, and orchestrator. The canonical pattern is ReAct — Thought, Action, Observation — which forces explicit reasoning before action. Key production concerns are state management, tool selection accuracy, cost and latency unpredictability, and observability. The main failure modes are infinite loops, hallucinated tools or arguments, context bloat, and cascading errors — mitigated by iteration caps, schema validation, summarization, and tracing.