𝗙𝗿𝗼𝗺 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝘁𝗼 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀: 𝗧𝘄𝗼 𝗬𝗲𝗮𝗿𝘀 𝗮𝘀 𝗮𝗻 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿
Two years ago, I only used AI to ask questions.
Today, I orchestrate multiple coding agents. I connect company knowledge through MCP. I run local models in iOS apps. I maintain a memory layer so agents can work together.
I am not an AI researcher. I am an ordinary engineer who kept experimenting.
Here is my journey from chat to agents.
𝗦𝘁𝗮𝗴𝗲 𝟭: 𝗧𝗵𝗲 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗚𝗮𝗽 At first, AI felt like magic. Then, it felt unreliable. Models often present wrong information with high confidence. I learned a hard lesson: A plausible answer is not a reliable result.
𝗦𝘁𝗮𝗴𝗲 𝟮: 𝗔𝗜 𝗮𝘀 𝗮 𝗣𝗮𝗿𝘁𝗻𝗲𝗿 Tools like Cursor changed everything. AI moved from a separate chat window into my code editor. The feedback loop became incredibly short: • Describe an idea. • Generate code. • Run and observe failure. • Ask for a fix. • Repeat.
This lowered the cost of building prototypes. Small ideas no longer die during setup or configuration. They actually reach the finish line.
𝗜 𝗹𝗲𝗮𝗿𝗻𝗲𝗻𝗱 𝘁𝗵𝗮𝘁 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝘀 𝗞𝗲𝘆 Early on, agents would forget decisions or break architecture. I stopped focusing on "prompt engineering" and started "context engineering." I began writing strict rules for my agents: • Follow the existing architecture. • Explain the plan before changing code. • Do not delete files without asking.
We were not just writing better sentences. We were building a stable environment for a probabilistic system.
𝗦𝘁𝗮𝗴𝗲 𝟯: 𝗕𝗲𝘆𝗼𝗻𝗱 𝘁𝗵𝗲 𝗖𝗵𝗮𝘁 𝗪𝗶𝗻𝗱𝗼𝘄 I moved from web chats to APIs and local models. Running a model on a desktop is easy. Shipping a model inside a mobile app is hard. Suddenly, I had to solve real engineering problems: • Model size and memory usage. • Startup latency. • Offline execution and device compatibility.
𝗦𝘁𝗮𝗴𝗲 𝟰: 𝗔𝗴𝗲𝗻𝘁-𝗗𝗿𝗶𝘃𝗲𝗻 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 The focus shifted from "write this function" to "understand this repository and complete this goal." I also started using the Model Context Protocol (MCP). I built an MCP integration for my company's knowledge platform. The agent could now access ten years of company data. The challenge moved from model intelligence to system design.
আমার সবচেয়ে বড় শিক্ষাগুলো: • কোডিং সহজ, কিন্তু ইঞ্জিনিয়ারিং কঠিন। • AI ইমপ্লিমেন্টেশন সামলায়, কিন্তু বিচারবুদ্ধির কাজ আপনাকে করতে হবে। • আপনাকে সুনির্দিষ্টভাবে লক্ষ্য নির্ধারণ করতে হবে এবং ধারণাগুলো যাচাই করতে হবে। • মাল্টি-এজেন্ট কাজের জন্য একটি মেমরি লেয়ার প্রয়োজন যাতে এজেন্টরা পুনরায় শুরু না করেই কাজ হস্তান্তর করতে পারে।
AI ইমপ্লিমেন্টেশনকে সস্তা করে তুলেছে। এর মানে হলো পরীক্ষা-নিরীক্ষা এখন বিনামূল্যে করা সম্ভব। আসল কাজ হলো কোন ধারণাগুলো আসলে তৈরি করার মতো যোগ্য তা সিদ্ধান্ত নেওয়া।
ঐচ্ছিক লার্নিং কমিউনিটি: https://t.me/GyaanSetuAi