Agents

An agent is an LLM that doesn't just produce text — it takes actions in a loop, observes the results, and decides what to do next. Give a model tools (search a database, call an API, run code, read a file) and a goal, and it can chain steps toward that goal rather than answering in one shot.

This is powerful and also where things get genuinely risky: an agent that can act can act wrongly, and it processes untrusted content that may try to hijack it. So this section is as much about architecture and guardrails as about capability.

Documents in this section

• ReAct and Tool Use — the reason-act-observe loop that underlies most agents.
• Agent Architectures — from a single tool-using loop to planners and multi-agent systems, and when to use which.
• Planning and Memory — how agents decompose tasks and carry state across steps.
• Guardrails for Agents — keeping an acting model safe, bounded, and debuggable.

Prerequisites: structured output / tool schemas (agents act by emitting tool calls) and the Safety section (agents are the highest-stakes place prompt injection shows up).