Agentic AI (AI Agents)

Agentic AI désigne les systèmes IA autonomes utilisant LLMs comme orchestrateur central pour planifier, décomposer, exécuter et adapter des plans pour accomplir des tâches multi-étapes. Évolution clé 2023-2025 — du LLM "text generator passif" à l'agent qui agit dans le monde via tools, code execution, web browsing, file manipulation.

Capabilities typiques d'agents :
(1) **Tool use / Function calling** — LLM décide quels outils invoquer (search web, query DB, send email, run code) basé sur task.
(2) **Planning** — décomposer goal complexe en sub-tasks, ordering.
(3) **Memory** — short-term (conversation), long-term (vector DB persistent).
(4) **Reflection** — review own actions, self-critique, replan if needed.
(5) **Multi-agent collaboration** — multiple agents specialized roles (researcher, writer, coder, critic) collaborent.
(6) **Code execution** — write et run code sandbox pour data analysis, calculations.
(7) **Web browsing** — navigate websites, click buttons, fill forms.
(8) **Computer use** — control desktop via screenshots + mouse/keyboard (Anthropic Claude Computer Use 2024).

Frameworks populaires :
(1) **LangChain** + **LangGraph** — agent orchestration with graphs ;
(2) **LlamaIndex** Agent Workflows ;
(3) **CrewAI** — multi-agent crews specialized roles ;
(4) **AutoGen** (Microsoft Research) — multi-agent conversation framework ;
(5) **OpenAI Swarm** (Oct 2024) — lightweight multi-agent ;
(6) **OpenAI Assistants API** ;
(7) **Anthropic SDK** with tool use natif et MCP (Model Context Protocol) ;
(8) **smolagents** (HuggingFace) — code agents ;
(9) **Phidata** ;
(10) **Pydantic AI** ;
(11) **DSPy** for declarative agent programming.

Use cases production : (1) coding agents (GitHub Copilot Workspace, Cursor, Devin, Claude Code, Replit Agent) — autonomous code modifications, multi-file refactors ; (2) research agents (Perplexity, Bing Chat) — multi-source web research ; (3) data analyst agents — query DBs, generate insights ; (4) customer support agents ; (5) sales prospecting agents ; (6) browser automation (Multi-On, Browser Use) ; (7) workflow automation (Zapier AI Actions, n8n) ; (8) game NPCs.

Défis : (1) **Reliability** — agents fail unpredictably mid-task, hallucinations ; (2) **Cost** — many LLM calls per task ; (3) **Safety** — autonomous agents avec accès filesystem/email/credit cards = potentiel disaster ; (4) **Evaluation** — hard to test (open-ended outputs) ; (5) **Long horizon planning** — current LLMs struggle 50+ steps. Compétences AI-102, AIF-C01.