# Generative AI and Agentic AI: From Answering Questions to Taking Action

> Source: <https://dev.to/agenticdevops/generative-ai-and-agentic-ai-from-answering-questions-to-taking-action-2il9>
> Published: 2026-06-05 13:12:03+00:00

*Pipeline & Prompts | Byte size guides on DevOps, Cloud and AI*

I used to spend hours before important workshops and client meetings pulling together materials. Downloading presentation decks, saving links, reading through documents, trying to hold enough in my head to answer questions confidently in the room.

Then I discovered NotebookLM.

NotebookLM is a generative AI tool from Google that lets you upload your own documents, presentations, and links and then have a conversation with all of that material as if it were a knowledgeable colleague who had read everything perfectly.

Before a recent conference I uploaded every presentation, every link, every document related to the topics being discussed. Then during the event, when someone asked me a question I was not sure about, I could query my own private knowledge base in seconds and get a precise answer drawn directly from the materials I had loaded.

I also started using generative AI to write meeting agendas, workshop plans, and to evaluate presentation slides before I delivered them — asking the AI to critique the flow, identify gaps, and suggest improvements the way a trusted colleague might.

The difference in how I work before and after generative AI is not subtle. It is significant.

And I am only just getting started with what comes next.

In Article 8 we covered the four types of AI — descriptive, predictive, prescriptive, and generative — using Google Maps as our guide. We talked about how generative AI like ChatGPT and Claude can create new content — text, code, images — rather than just analysing existing data.

In this article we are going deeper into two things:

A Large Language Model — or LLM — is the technology behind ChatGPT, Claude, Gemini, and most of the generative AI tools you encounter today.

Here is how to think about how it was built.

Imagine reading every book, every article, every website, every piece of code ever written — billions and billions of words. As you read, you start to notice patterns. Certain words appear together. Certain sentence structures follow certain ideas. Certain concepts are always explained in similar ways. Over time you develop an incredibly deep intuition for language — not because you memorised everything, but because you absorbed the patterns.

That is essentially what happens when an LLM is trained. It processes an almost incomprehensibly large amount of text and learns the statistical patterns of language — which words follow which words, which ideas connect to which concepts, how questions are typically answered.

When you ask it a question it does not look up the answer in a database. It generates a response word by word, each word chosen based on what is most likely to come next given everything it has learned. This is why it feels like a natural conversation — because it is built on the patterns of natural human communication.

It is also why hallucinations happen. The model is always generating what is most likely — not necessarily what is true. If the patterns in its training data point toward a plausible but incorrect answer, that is what it produces with complete confidence.

Understanding this does not make LLMs less useful. It makes you a better user of them.

**Drafting and communication**

First drafts of emails, reports, agendas, presentations, and proposals. Generative AI is extraordinarily fast at producing a solid starting point that you then refine. The key word is starting point — your judgment, your knowledge, and your voice still matter.

**Explaining complex things simply**

Ask an LLM to explain a concept as if you are a complete beginner and it will almost always produce a clearer explanation than most textbooks. This is one of its most underrated uses — using it as a patient, infinitely available teacher.

**Code review and debugging**

As I shared in Article 8, using AI to review Terraform code as a beginner was genuinely transformative. It applies the pattern recognition of someone who has seen thousands of codebases to your specific code. GitHub Copilot does this in real time as you type.

**Summarising and synthesising**

Upload a long document, a set of meeting notes, or a collection of links and ask for a summary. NotebookLM does this particularly well because it works only with the materials you provide — so the answers are grounded in your specific content rather than general training data.

**Brainstorming and thinking partners**

Some of the most valuable AI interactions are not about getting an answer but about thinking through a problem. Describe a challenge you are facing, ask the AI to push back on your assumptions, and you often end up with a clearer view of the problem than when you started.

Since we keep it honest on this blog:

**Hallucinations are real.** AI models can generate confident, detailed, completely wrong answers. Always verify important facts, especially anything involving numbers, dates, specific technical details, or legal and medical information. Critical thinking is not optional.

**It does not know what it does not know.** An LLM will rarely say “I have no idea.” It will generate something that sounds plausible. The less common or more specialised your question, the more carefully you need to evaluate the response.

**Context has limits.** Most LLMs can only hold a certain amount of conversation in their working memory at once. Very long or complex projects can lose earlier context in ways that affect the quality of later responses.

**It has no real world awareness.** A standard LLM does not know what happened yesterday, cannot access your company’s internal systems, and cannot take action in the world on your behalf.

That last limitation is exactly what Agentic AI is designed to solve.

Every AI tool we have discussed so far follows the same basic pattern. You ask a question. The AI answers. You ask another question. It answers again. Each interaction is a single exchange.

**Agentic AI breaks that pattern.**

An AI agent does not just answer a question — it takes a goal, breaks it down into steps, uses tools and external systems to complete those steps, learns from what it finds along the way, and delivers a completed result.

Think of the difference between asking a colleague a question and delegating a task to them.

Asking a question: “What flights are available from London to New York next Tuesday?”

Delegating a task: “Book me the best value flight from London to New York next Tuesday, add it to my calendar, and email the confirmation to my team.”

The first is what current generative AI does. The second is what agentic AI does — or is rapidly learning to do.

I am currently experimenting with **n8n** — an open source workflow automation tool that lets you connect AI to real world actions and systems.

With n8n you can build workflows where AI is not just thinking but doing. For example:

Here is a simple example of what an n8n workflow looks like conceptually:

```
Trigger: New email arrives
   ↓
Step 1: AI reads the email and classifies it
   ↓
Step 2: If it contains an action item → create a task
Step 3: If it is a meeting request → check calendar and respond
Step 4: If it is a document → summarise and file it
   ↓
Result: Inbox managed automatically
```

I am still in the early stages of building with n8n — it is a genuine work in progress. But even at this early stage the potential is clear. Tasks that used to require human attention for every individual item can be handled automatically, consistently, and at a scale no individual could match.

It helps to think of AI capability as a spectrum rather than a binary on/off:

**Level 1 — AI Assistant**

Answers questions, generates content, reviews code. You drive every interaction. ChatGPT and Claude in basic use are here.

**Level 2 — AI with Tools**

The AI can search the web, read documents, run code, and access external data to give better answers. Claude with web search enabled, NotebookLM, and GitHub Copilot are here.

**Level 3 — AI Workflow Automation**

The AI completes multi step tasks using connected tools and systems. n8n workflows with AI, Zapier with AI integration, and similar tools are here. This is where I am currently experimenting.

**Level 4 — Fully Autonomous AI Agents**

The AI receives a high level goal and figures out how to achieve it independently — planning, executing, adapting, and reporting back. This is where the industry is heading and where the most significant questions about oversight and control are being asked.

We are moving along this spectrum faster than most people realise.

In Cloud and Infrastructure work, agentic AI is beginning to appear in ways that will fundamentally change the role of the DevOps engineer.

**Automated incident response** — AI agents that monitor your Kubernetes cluster, detect an anomaly, diagnose the likely cause, apply a fix, and log the resolution — without waking anyone up at 3am.

**Self healing infrastructure** — Terraform configurations that AI can review, update, and apply in response to changing requirements or detected drift — the configuration drift problem we talked about in Article 7, solved autonomously.

**Intelligent CI/CD pipelines** — Pipelines that do not just run fixed steps but adapt based on what they find. If a test fails, the agent analyses why, suggests a fix, and in some cases applies it automatically.

**AI assisted cloud cost optimisation** — Agents that continuously monitor your cloud spend, identify wasteful resources, and make recommendations or take action to reduce costs without manual review.

None of this is fully here yet in the way the marketing materials sometimes suggest. But the direction is clear and the pace of change is faster than anything I have seen in a decade of working in Cloud and Infrastructure.

As AI becomes more capable of taking action — not just giving advice — the most important skill is not knowing how to use AI. It is knowing when to trust it, when to verify it, and when to keep a human in the loop.

The engineers and professionals who will thrive in an agentic AI world are not the ones who hand everything to the AI and walk away. They are the ones who understand what the AI is doing well enough to catch it when it goes wrong and to guide it toward genuinely useful outcomes.

That combination of AI capability and human judgment is going to be the most valuable professional skill of the next decade.

Here is everything we covered today:

We have now covered the full Pipeline & Prompts foundation — from DevOps and Linux all the way through to Generative and Agentic AI.

In Article 10 we are going to zoom out and look at the **big picture — how DevOps, Cloud, and AI are converging into a single discipline** and what that means for anyone building a career in technology right now.

We will also talk about where to go next on your learning journey — the resources, the certifications, and the hands on projects that will take you from understanding these concepts to actually building with them.

The foundation is laid. Now it is time to build.

See you in Article 10.

*Written by Pipeline & Prompts | Byte size guides on DevOps, Cloud and AI*

*Found this useful? Share it with someone who thinks AI is just a chatbot — and show them how much further it has already come. Follow along for a new article every week.*