Tech, Rust, Cybersecurity, and Notes Zola 2025-07-19T00:00:00+00:00 https://blogs.alokranjan.me/tags/llm/atom.xml 2025-07-19T00:00:00+00:00 2025-07-19T00:00:00+00:00 ryu https://blogs.alokranjan.me/blog/agentic-ai/ <h2 id="what-is-an-agentic-ai">What is an Agentic AI</h2> <p><strong>Agentic AI</strong> refers to systems that <em>don’t</em> just answer questions, but can plan, decide, and act on their own.</p> <p>They use LLMs + Tools + Memory to handle Multi-Step Tasks without being told exactly what to do at each step.</p> <h3 id="for-example">For example:</h3> <ul> <li>You say “<strong>Book me a trip to Delhi this weekend under Rs.5000.</strong>” <ul> <li> <p>An <em><strong>Agentic AI</strong> might</em>:</p> <ul> <li>Search for flights &amp; buses.</li> <li>Check Hotels.</li> <li>Compare Costs.</li> <li>Make bookings via APIs.</li> <li>Email you the plan.</li> </ul> </li> <li> <p>Now, this is way beyond just generating text - this is actual <strong>Decision + Action</strong>.</p> </li> </ul> </li> </ul> <hr /> <h2 id="characterstics-of-agentic-ai">Characterstics of Agentic AI:</h2> <ol> <li> <p><strong>Goal-Oriented Planning</strong></p> <ul> <li>It starts <strong>with a goal</strong>, then breaks it into smaller steps and executes.</li> </ul> </li> <li> <p><strong>Multi-Step Reasoning</strong></p> <ul> <li>It doesn’t stop at one reply - it <em>iteratively</em> reasons until the goal is met.</li> </ul> </li> <li> <p><strong>Autonomous Decision-Making</strong></p> <ul> <li>It chooses <strong>the best path</strong> or the tool on its own, based on the whatever context it has in that situation.</li> </ul> </li> <li> <p><strong>Uses Memory, Tools and Knowledge</strong></p> <ul> <li>Remembers Past Events (long term memory or conversational memory), uses APIs/Scripts (tools), and refers to the docs or rules (Knowledge Base).</li> </ul> </li> </ol> <hr /> <h2 id="what-makes-agentic-ai-different-from-basic-llms">What makes Agentic AI different from Basic LLMs?</h2> <ul> <li>Normal LLMs (Chat GPT, Gemini, Perplexity etc.) are <strong>Passive</strong>: <ul> <li>They just respond to a prompt.</li> </ul> </li> <li>Agentic Systems are <strong>Active</strong>: <ul> <li>They observe, think, plan, act, and even revise their actions.</li> </ul> </li> </ul> <hr /> <h2 id="basic-agentic-structure">Basic Agentic Structure</h2> <figure> <p><img src="https://blogs.alokranjan.me/blog/agentic-ai/agentic.png" alt="agentic-ai" /></p> <figcaption>Architecture of an Agentic AI</figcaption> </figure> <br /> <hr /> <h2 id="agentic-ai-vs-ai-workflow">Agentic AI VS AI Workflow</h2> <table><thead><tr><th>Term</th><th>What it means</th></tr></thead><tbody> <tr><td>AI workflow</td><td>Pre-defined steps using AI (e.g. RAG pipeline)</td></tr> <tr><td>Agent</td><td>Dynamic reasoning &amp; acting system that <strong>adapts</strong></td></tr> <tr><td>ReAct Agent</td><td><em>Reason + Act</em>: Think, Decide, take Action</td></tr> <tr><td>Control logic</td><td>Like if/else to guide actions</td></tr> <tr><td>Tools</td><td>External Functions/APIs</td></tr> <tr><td>Memory</td><td>Past info/context retained</td></tr> <tr><td>Planner</td><td>Creates a step-wise plan</td></tr> </tbody></table> <hr /> <h2 id="llms-in-this-context">LLMs in this context</h2> <ul> <li><strong>LLMs</strong> (like GPT, Gemini, Claude) are the <strong>Core Brains</strong> of the system.</li> <li>But on their own, they can’t remember, act or plan.</li> <li>We wrap them in <strong>Agents</strong> to extend their <em>ability</em>.</li> </ul> <hr /> <h2 id="agentic-ai-example">Agentic AI Example</h2> <ol> <li><strong>AI Coding Assistant</strong> <ul> <li>Can Understand bug report, edits code, runs tests, commits.</li> </ul> </li> <li><strong>Travel Agent</strong> <ul> <li>As we have already seen the example above; There are a <em>gazillion</em> use-cases of Agentic AI in various domains.</li> </ul> </li> <li>A few more examples would be an <em>HR Onboarding Bot</em>, An Agent that posts for you on LinkedIn, the list <em>goes on</em>~</li> </ol> <hr /> <h2 id="let-s-recap">Let’s recap</h2> <ul> <li>LLM = <strong>Brain</strong></li> <li>Agent = <strong>Brain + Body</strong>(Tools/Memory/Planner)</li> <li>Agentic AI = <strong>Autonomous Worker</strong> that thinks and achieves <strong><em>any goal</em></strong> given to it</li> <li>Multi-Turn Planning = Not just simple answers, but <em>strategic answers</em> driven by context</li> <li>ReAct Loop = <strong>Reason -&gt; Act -&gt; Observe -&gt; Repeat</strong>.</li> </ul> </br> <blockquote class="important"> <p class="alert-title"> <i class="icon"></i>Important</p> <p>Before finishing up this blog; I would like for you to read this <a href="https://www.anthropic.com/engineering/building-effective-agents">Amazing Blog by Anthropic on building Agentic AI</a>.</p> </blockquote>