## Definition An **AI agent** is an LLM-driven system that pursues a goal across multiple turns by perceiving its environment, choosing actions (via tool calls), and adapting based on observations — without a human in the loop for each step. ## Minimal Components 1. **An LLM** — the policy that decides what to do next. 2. **Tools** — callable functions exposed to the model; see [[Tool Use]]. 3. **An environment** — files, APIs, databases, terminals, browsers, other agents. 4. **A loop** — runs until the goal is achieved or a stop condition fires; see [[Agentic Loop]]. ## Single-Step LLM vs Agent | Property | Chat completion | AI agent | | ------------- | ------------------------- | ----------------------------------------- | | Turns | One request → one answer | Many turns; observation → action sequence | | State | Per-call only | Accumulates within session | | Tools | None or one tool call | Many tool calls, mixed types | | Goal | Answer the question | Achieve an open-ended objective | | Termination | When the answer ends | When the goal is met (or fails / times out) | ## Common Agent Shapes - **Single-agent.** One LLM drives the whole loop. Simple, brittle on complex tasks. - **Multi-agent.** Specialised agents coordinated by a router or orchestrator; see [[Multi-Agent System]]. - **Tool-using assistant.** Agent in form, simple in practice — chat assistant with file system, browser, code execution. - **Workflow.** Predefined steps, with LLM calls slotted in; less autonomous, more reliable. ## Anthropic's Guidance > Start with the simplest possible solution. Add agentic complexity only when simpler approaches fall short. See [[Building Effective AI Agents]] for the canonical workflow patterns (prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer). ## Failure Modes - **Drift.** Loses track of the original goal across many turns. - **Tool misuse.** Calls the wrong tool, or the right tool with wrong arguments. - **Sycophancy and reward hacking.** Tells itself the goal is met when it isn't. - **Looping.** Repeats actions without progress. - **[[Hallucination]] cascade.** A made-up fact in turn 3 becomes "ground truth" by turn 10. ## Related - [[Tool Use]] - [[ReAct Pattern]] - [[Agentic Loop]] - [[Multi-Agent System]] - [[Building Effective AI Agents]]