IA generativa gera. IA agentic age

Generative AI and agent AI, both use the same model, but it's not in the README. The difference lies in what they do. You've likely encountered a chat where the "agent" resolved everything in Markdown, and the git repository remained unchanged? That's exactly what happened during my Introduction to Agent AI module 1 of Agentic AI with LangChain and LangGraph on Coursera (IBM).

The LangGraph versus LangChain debate became dense with chain, graph, and harness. Here, it's simpler: generative AI takes a prompt, outputs text – that's it. Next word. Translation, summarization, suggestion of a Markdown patch, and it's done. No git history, no database, no gateway interaction.

Agent AI, on the other hand, has the model choose the next step in the world. It reads a file, calls an API, opens a pull request, starts a payment. There's a loop, there's a tool, there's a consequence. What you take from this: generating is text. Acting is tool usage. Look at the history, not the README. One fact: generating isn't acting.

Good generative AI is already valuable. Issue classification, test sketch, diff explanation – all done by generative AI. Calling this an "agent" is incorrect. You've seen a README with "AI agent" and code that's just an invoke. The agent here is the loop: observe, decide, use tool, see result, decide again. Pedagogical scene (not production logs): you cancel subscription 441, AI generates "the path cost is DELETE /v1/subscriptions/{id}", acts with an HTTP tool, gets a 204 response.

The AI subscription is canceled. The first is a hypothesis. The second is the world. Without a file or line in a review, we've learned not to mix finding with guesswork. Here, the same muscle is at work: text ≠ effect.

To change the world, a loop that changes the world needs a person in the middle. Reading a README is a tool, git push is another. Least privilege: both don't run on the same conveyor belt. HITL (Human in the Loop) where something is recorded. Blast radius: if the agent errs, the worst case is a bad sentence or a deployment? That's AppSec for agents, not slide presentation.

The dense chain, graph, and harness enter the Python runtime. This note: have you let the model act? A sandbox, allowlist, audit trail: the turn that only generates doesn't need them. The turn that pays does. The harness decides, not the vocabulary. Cursor, Claude Code, Codex, a StateGraph in the API: all can be generative in one turn and agent-like in another.

Look at the history. It's text versus tool that reads versus tool that writes. Only completed Markdown? Generative. Called a search, Read, gh? Generative tool. Called merge, migrate, deploy? Agent that writes. No human in the middle, incident. RAG / MCP / fine-tuning layers are another story. Chain, graph, and harness: the dense sibling, with the prompt to paste.

Here, the knob is either generate or act. In the final chat: they only wrote, read some tool, or wrote in the world?