AI agents: from chatbot to autonomous actor

This is the move from the copilot (which assists you while you work) to the digital worker (to whom you delegate the entire task). This shift is so central that the 2025-2026 period was widely dubbed "the year of agents."

In concrete terms, the process unfolds in two stages. Beforehand, the documents are split into pieces (chunks) and each is turned into an embedding (Chapter 2), a vector of numbers that captures meaning, stored in a vector store. At question time, the question too is converted into a vector, the pieces whose meaning is closest are retrieved (semantic search) and added to the prompt. The benefits are threefold: up-to-date and specialized answers (on private data the model has never seen), fewer hallucinations, and the ability to cite its sources, and therefore to be verified. It is today the flagship building block of enterprise applications.

Classic RAG retrieves once, then answers. Agentic RAG goes further by entrusting retrieval to an agent: it decides whether to search, reformulates the query, queries several sources or tools, assesses the quality of what it has found, and starts over if it is insufficient, before synthesizing. Where simple RAG is a reflex, agentic RAG is a small investigation: it adapts to complex, multi-step questions, at the cost of higher expense and latency. It is one of the ways in which the line between "a model that answers" and "an agent that acts" blurs.

Two projects dominate this category, with contrasting profiles: OpenClaw, the viral pioneer, and Hermes Agent, more safety-minded. They are significant enough, and representative enough, to each merit a case study (sections 6.8 and 6.9). Together, they illustrate both the democratization of agents (sovereignty, local data, Chapter 9) and the risks specific to highly autonomous agents (Chapter 20).

This image of "AI corporations" working in concert is no trifling matter: it is precisely the vision described by the most advanced prospective scenarios, in which thousands of copies of a model collaborate at a superhuman pace. We will return to it in Chapter 24, for it lies at the heart of the questions of alignment and control.

---

Key takeaways (Chapter 6)

• An agent does not answer, it acts: you entrust it with a goal and it accomplishes it autonomously, through a "think, act, observe" loop. It is the move from the copilot to the digital worker.
• An agent combines a brain (LLM) with four faculties: perception, memory, tools, planning.
• MCP (an open standard launched by Anthropic in late 2024, adopted by the industry) is "the USB-C of AI": it connects any AI to any tool.
• Frameworks (LangChain/LangGraph, CrewAI, AutoGen) and no-code tools (n8n, Make, Zapier) make building agents easier.
• Computer-use agents use a computer like a human, which is powerful but raises risks of reliability and security (prompt injection).
• Multi-agent systems get several agents to collaborate, prefiguring the "AI corporations" of prospective scenarios (Chapter 24).
• A wave of self-hosted personal agents (OpenClaw, Hermes Agent) runs an assistant that acts on your own machine, model-agnostic and with persistent memory: strong autonomy and sovereignty, but a new attack surface (Chapter 20).
• Vibe coding (describing software in natural language and letting the AI write it) democratizes software creation and accelerates experts, but exposes you to poorly understood code, security flaws (Chapter 20), and deskilling.

We have surveyed the "what" and the "how." Chapter 7 draws the map of the "who": the American, Chinese, and European labs, the divide between closed and open models, and the faces of the people shaping this revolution.

---