Agentic AI
LLM-system som tar autonomous-actions: planerar multi-step task, anropar tools, observerar resultat, anpassar plan. Skiljs från chat (en input → en output). Mainstream 2024 efter ChatGPT Plugins, Claude Tool Use, GPT Assistants.
Komponenter: LLM (resoning), tools (function-calling), memory (conversation/scratchpad), planner. Frameworks: LangChain, LlamaIndex, AutoGPT (2023-hype), Anthropic SDK, OpenAI Agents SDK. Use cases: coding-agents (Cursor, Aider, Claude Code), research-agents (Perplexity Deep Research, OpenAI Operator), e-commerce-agents (Amazon Q). Risks: prompt injection, runaway-actions, hallucinated tool-calls. Modern best practice: structured-output (JSON schema-validation per step), explicit-tool-confirmation för destructive actions, sandboxed-execution.