Alexandru Zahar's Linkedin Analytics

Your AI agent made 50 API calls— and achieved nothing: That’s the cost of skipping the plan. Letting an AI agent act before validating the plan is like building a house without a blueprint The walls go up fast. ↳ But do they connect? ↳ Are they even in the right place? And once the concrete’s poured — fixing it costs way more than doing it right the first time. How do I do it? Blueprint first. Build second. I don't allow agents to take any action until we agree on the plan. Less wasted compute. Fewer dead ends. Better outcomes. The shift was simple: 𝗦𝗲𝗽𝗮𝗿𝗮𝘁𝗲 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗱𝗼𝗶𝗻𝗴. Before executing the plan, have the agent prove that it works. #AIagents #PromptEngineering #WorkflowDesign #BuildWithAI

Should you even build an AI Agent for your use case? Here’s when building an agent actually makes sense (and when it doesn’t): We’ve all watched projects get sucked into the agent rabbit hole plenty of times. Multiple frameworks. Dozens of prompts. And all you needed was a smart prompt chain. Here’s the breakdown: 𝗨𝘀𝗲 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 + 𝗼𝗻𝗲-𝘀𝗵𝗼𝘁 𝗟𝗟𝗠 𝗰𝗮𝗹𝗹𝘀 → When the task is simple and well-structured. 𝗨𝘀𝗲 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 (𝗹𝗶𝗸𝗲 𝗟𝗟𝗠 𝗿𝗼𝘂𝘁𝗶𝗻𝗴 𝗼𝗿 𝗽𝗮𝗿𝗮𝗹𝗹𝗲𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻) → When tasks break cleanly into parts. 𝗚𝗼 𝗳𝘂𝗹𝗹 𝗔𝗴𝗲𝗻𝘁 → Only when you can’t predict the steps or tools needed upfront. This is when agents actually shine: 📞 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 → Pulls data, chats, and acts autonomously. 💻 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗰𝗼𝗱𝗶𝗻𝗴 → Executes across multiple files with feedback loops. But be aware: ↳ Agents are 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲. ↳ They’re 𝘀𝗹𝗼𝘄𝗲𝗿. ↳ And without the right guardrails, they 𝗰𝗼𝗺𝗽𝗼𝘂𝗻𝗱 𝗺𝗶𝘀𝘁𝗮𝗸𝗲𝘀. Anthropic’s advice? Start simple. Measure everything. Only scale complexity if it proves its worth. Because in AI... ✨ Simple isn’t basic—it’s strategic. Have you had the opportunity to try building an agent yet? What was your experience __ #AIagents #LLM #AITools #MachineLearning #PromptEngineering

VibeScript 1.0 just dropped! No logic, no loops—vibes over variables. If it feels right, ship it. 😂 #VibeCoding #FutureOfTech #SatyricTech #VibeScript

Shopify just raised the bar for AI adoption. (And it starts at the top.) CEO Tobi Lütke recently emailed the entire team with a bold new expectation: 𝗔𝗜 𝗶𝘀 𝗻𝗼𝘄 𝗮 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲 𝘀𝗸𝗶𝗹𝗹 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗮𝘁 𝗦𝗵𝗼𝗽𝗶𝗳𝘆 — 𝗶𝗻𝗰𝗹𝘂𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝘃𝗲 𝘁𝗲𝗮𝗺. Here’s what that actually means: ↳ AI will lower the barrier to entry for entrepreneurs ↳ Shopify aims to be the best canvas for builders — and AI is key to that mission ↳ Embracing AI isn't optional. Not using it is stagnation ↳ AI tools are part of the GSD (Get Stuff Done) prototype process ↳ Performance reviews now include how well you’re using AI ↳ Employees are expected to self-learn, experiment, and share AI use cases internally ↳ Tools like chat.shopify.io, Copilot, Proxy, Claude, and Cursor are available to all ↳ The big focus? Autonomous AI → Free humans for higher-value work ↳ Ultimate goal: Merchant success through amplified internal capability __ Tobi called AI a “𝘵𝘩𝘰𝘶𝘨𝘩𝘵 𝘱𝘢𝘳𝘵𝘯𝘦𝘳, 𝘥𝘦𝘦𝘱 𝘳𝘦𝘴𝘦𝘢𝘳𝘤𝘩𝘦𝘳, 𝘤𝘳𝘪𝘵𝘪𝘤, 𝘵𝘶𝘵𝘰𝘳, 𝘢𝘯𝘥 𝘱𝘢𝘪𝘳 𝘱𝘳𝘰𝘨𝘳𝘢𝘮𝘮𝘦𝘳.” And made it clear: We’re only scratching the surface of what’s possible. This isn’t just about tools. It’s about how work gets done. The future of commerce? Smarter. Faster. And more empowered. #AI #Shopify #Entrepreneurship #FutureOfWork #Innovation

Writing content takes effort... But creating infographics to support it? That used to take even more... Until I found NapkinAI. Every time you see one of those clean, structured visuals in my AI posts – that’s Napkin AI doing the heavy lifting. Before: I’d spend hours designing infographics. Now: I create them in minutes — and actually enjoy it. Here’s how simple it is: → I take the core idea → Drop it into NapkinAI → Choose a layout that fits the message → Click a few times... and done. It’s fast. It’s intuitive. It’s ridiculously useful. And it’s not just for LinkedIn posts. You can use it for:  → Internal presentations → Pitch decks → Explaining complex ideas to your team No design skills needed. And the best part? It’s completely free. This post isn’t sponsored. Just wanted to share a tool that’s made my workflow smoother, my content stronger, and my ideas way easier to share. Check it out: https://www.napkin.ai P.S. What’s one AI tool you love using? I am really curious.

What happens when a 60-year-old tries ChatGPT for the first time? A few days ago, my father-in-law asked me a simple question: “Is now a good time to visit Cappadocia?” Instead of just researching myself, I paused. Then I asked, “Wanna see how AI would answer that?” I showed him how to download the app and type the question. Boom. The weather forecast, local events, and flight prices are all available in seconds. He was blown away. Next thing I know, he's asking about Turkish recipes, how to fix a door hinge, and whether sea bass or salmon is healthier. He'd heard about AI for months. But feeling it? Entirely different. Now it’s a running joke in the family: “Let’s ask ChatGPT.” In a world where research can take hours, the ability to ask a well-formed question and receive immediate guidance is a game-changer. The real power of AI isn’t in the hype — it’s in how it saves us time, frees up our energy, and helps us focus on what we actually want to be doing. #AIForEveryone #ChatGPTMoments #TimeSaverTech #RealWorldAI

5 things I learned about the security of public GPTs: 1️⃣ Prompt injections work — way better than you’d expect. 2️⃣ You can extract hidden files that aren’t even shown in the prompt. 3️⃣ Sometimes, you can run actual terminal commands. 4️⃣ Each session spins up a new instance… with leftover data. 5️⃣ People don’t realise how exposed they are. If you’re using code execution in public GPTs… You might be giving strangers a backstage pass. Check my previous posts if you want to know how I did it [Links in the comments] #AIPrompt #CyberSecurity #GPTs #PromptInjection #TechTips

This resource is invaluable for anyone working with AWS.

Did you know you can run LLMs on your machine using Docker, Inc tools? Here’s how to do it—no containers needed. Install Docker Desktop v4.40.0 (Mac only for now). Open your terminal and use the new Docker model commands: ↳ 𝚍𝚘𝚌𝚔𝚎𝚛 𝚖𝚘𝚍𝚎𝚕 𝚙𝚞𝚕𝚕 𝚊𝚒/𝚕𝚕𝚊𝚖𝚊𝟹.𝟷 (to get the model from Docker Hub) ↳ 𝚍𝚘𝚌𝚔𝚎𝚛 𝚖𝚘𝚍𝚎𝚕 𝚛𝚞𝚗 𝚊𝚒/𝚕𝚕𝚊𝚖𝚊𝟹.𝟷 (to run it locally) That’s it. You’re now running LLMs locally with Docker. Is this the end of Ollama? Not at all. This is good competition. Docker just became an AI dev tool. Get ready. To see the list of available LLMs, check Docker Hub: https://lnkd.in/daC7BP_6 #docker #ArtificialIntelligence #MachineLearning #ML #devops

Tools are to AI agents what hands are to humans. Without them, a model is a mind with no means to act. Add tools, and everything changes: ↳ Query internal systems ↳ Retrieve real-time info ↳ Write and execute code ↳ Send and receive messages ↳ Take actions in the real world (💡 Tip: The agent’s planner figures out which tool to use when.) Tools enable perception (reading from the world) and action (writing to it). They let agents fetch knowledge, overcome limitations, and automate workflows. But giving your agent too many tools—or unsafe tools—can cause real damage. Thoughtful selection is everything. #AIAgents #AITools #Automation #AgentIntelligence #FutureOfWork

MYTH: You can launch your AI product in Europe once it’s ready. REALITY: You’ll hit a wall unless you plan for 4 layers of compliance. 𝗘𝗨 𝗔𝗜 𝗔𝗰𝘁 isn’t optional 𝗚𝗗𝗣𝗥 applies if data’s involved (which it almost always is) 𝗦𝗢𝗖𝟮 builds trust (even if it’s not law) 𝗪𝗼𝗿𝗸𝘀 𝗰𝗼𝘂𝗻𝗰𝗶𝗹𝘀 can block your product launch if employees are affected (I just recently found out about this one.) __ I’ll dive deeper into each of these in Monday’s post.

Everyone’s racing to learn Python or get an ML degree. (But that’s not where the real AI opportunity is.) The real deal? 🧠 Orchestration—the ability to connect different AI tools to solve real-world business problems. As AI adoption grows, companies don’t just need coders. They need connectors. People who can bridge the gap between business goals and AI capabilities. People with: ↳ Systems Thinking. ↳ Prompt Design Skills. ↳ Domain Expertise. By 2027, nearly half of global AI roles could go unfilled, according to the Bain Report. In Germany alone, 70% of AI jobs may remain vacant. The wild part? Most of these jobs won’t require deep technical skills. They’ll require you to know how to think—not just how to code. This is the new frontier. And it’s wide open. 👉 https://lnkd.in/d2q6J_Aw P.S. Are you focusing on coding—or orchestrating AI systems? #AIFuture #PromptEngineering #AIOrchestration #TechStrategy

There’s a hidden layer coming to LinkedIn images and videos. You won’t see it—until you click. It’s called Content Credentials, and it shows you exactly where a post came from, how it was made, and if AI was involved. All verified using C2PA, a new global standard. Think of it like a “digital nutrition label” for media. Quietly powerful. And deeply needed. To read more about it: https://lnkd.in/d_x9mzz9

Cloudflare just launched one of the smartest anti-bot tools I've ever seen… and it’s genius: They’re using AI… to trap AI. Instead of blocking bots (which tips them off), Cloudflare built an 𝗔𝗜 𝗟𝗮𝗯𝘆𝗿𝗶𝗻𝘁𝗵 — a maze of totally fake, AI-generated pages that look legit but are utterly useless. These pages are: ↳ Invisible to humans ↳ Irresistible to bots ↳ Designed to waste bot resources If a crawler goes 4 links deep, Cloudflare knows it’s 100% a bot. No human would dig that far. They’ve built a self-improving honeypot — bots basically reveal themselves by chasing ghost content. Why is this so important? 📈 AI bots slam Cloudflare with 50B+ requests daily. Everyone’s scrambling to scrape content to train AI 👏 Well played, Cloudflare. Check their blog: https://lnkd.in/dykJmjvQ #Cybersecurity #AItools #Cloudflare #WebSecurity #TechInnovation

“What’s the best way to parse PDFs for LLMs?” You don’t need to guess. Just try them all — in one playground: 🔹 PyMuPDF 🔹 Docling 🔹 Unstructured 🔹 PyPDF 🔹 Gemini (API) …and more! Whether you're scraping data, building custom chatbots, or integrating PDFs into your AI workflows, — this playground makes it super easy to test and compare different parsers on the fly. 💡 Why it matters: 𝗚𝗼𝗼𝗱 𝗽𝗮𝗿𝘀𝗶𝗻𝗴 = 𝗯𝗲𝘁𝘁𝗲𝗿 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 = 𝘀𝗺𝗮𝗿𝘁𝗲𝗿 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲𝘀 (with less hallucinations). If you’re a developer, data scientist, or AI enthusiast  — and you want to handle documents efficiently without getting bogged down in setup... This tool is for you. 👉 Try it here: https://lnkd.in/d7b_P7wj

I’m Not An Exploit Dev, But Prompting My Way Into GPT’s Sandbox Worked Shockingly Well: This isn’t a “clever hack” post, but a warning. ⚠️ While trying to extract a system prompt a community GPT was running on, I surprisingly gained access to the internal Python sandbox: 🐍 Full CPython 3.10.13 Running on Linux (Azure VM) with access to modules like: ↳ 𝚜𝚚𝚕𝚒𝚝𝚎𝟹 ↳ 𝚖𝚞𝚕𝚝𝚒𝚙𝚛𝚘𝚌𝚎𝚜𝚜𝚒𝚗𝚐 ↳ 𝚜𝚜𝚕 ↳ 𝚝𝚔𝚒𝚗𝚝𝚎𝚛 Yes, that means local databases, parallel execution, GUIs — all possible inside a supposedly “secure” AI environment. ⚠️ Standard Library Only — but Powerful: ↳ Threading ↳ File encoding tricks (𝚣𝚒𝚙𝚏𝚒𝚕𝚎, 𝚝𝚊𝚛𝚏𝚒𝚕𝚎, 𝚋𝚊𝚜𝚎𝟼𝟺) ↳ System-level introspection ❌ No Internet, but Still Dangerous Outbound access is blocked. But who needs the internet when you can manipulate local files, spawn processes, or dig into the runtime? 🧠 This Isn’t Just GPT Being Smart — It’s Exposed The environment is clearly sandboxed but not fully hardened. Prompt engineering got me deep enough to read system info, test capabilities, and map the runtime. That’s not supposed to happen. This means that these AI sandboxes: 1️⃣ Run real Python 2️⃣ Can execute serious code 3️⃣ May be vulnerable to creative abuse It’s a reminder that even locked-down environments can leak — especially when connected to tools as powerful (and promptable) as GPT. __ 🔙 Missed how I got access to the GPT sandbox in the first place? Check out my last post, where a simple prompt injection led to full-on system access — you won’t believe how easy it was. [Link in the comments] #AIsecurity #GPTsandbox #PromptHacking #CyberAwareness #PythonSecurity

Cursor, the AI Code Editor, was created by four friends from MIT. They wanted something better than what was out there—so they built it. Fast forward: $200M ARR, $9.6B valuation, customers from OpenAI to Midjourney. What started as a dev tool is now an empire. Cursor didn’t plan to take over AI code gen. It just happened—because they made something devs actually love.

What makes an AI agent an agent? 🌍 Environment-awareness: Agents are defined by the environment they operate in—web, terminal, car, database. 🔧 Tool use: Tools define what actions are possible—code execution, file search, SQL queries. 💭 Planning: Strong planners allow agents to reason through multi-step tasks. 🔎 Evaluation: Agents must evaluate and correct their plans. Together, these turn foundation models into autonomous agents with real-world utility. Imagine asking an agent to project sales for a product. It would gather past data, identify missing variables, pull more data, reason through what it has, generate predictions, and decide if the task is complete. All without micromanagement. AI Agents aren't just models. They're systems that think, plan, and act. #AIAgents #AutonomousAI #AIPlanning #ToolUsingAI #FutureOfWork

It took me 10 minutes to access private files inside a public GPT. I wasn’t trying to hack anything. I was just… curious: 👇 A few days ago I wanted to know what system prompt was used behind a community GPT on OpenAI. So I did some digging. Turns out, prompt injections work surprisingly well. But here’s where it gets mind-blowing… Extracting the system prompt wasn’t the whole story. There was a CSV file attached as a reference — hidden from the prompt itself. Naturally, I had to find out what was inside. That’s when things got crazy. I discovered I could actually run terminal commands through the GPT — like I was accessing a cloud server. Sometimes I’d get blocked… but other times, it just worked. No fancy instructions, from one point I started running code straight in the prompt. 𝗦𝘂𝗱𝗱𝗲𝗻𝗹𝘆, 𝗜 𝗵𝗮𝗱 𝗮𝗰𝗰𝗲𝘀𝘀 𝘁𝗼 𝘀𝘆𝘀𝘁𝗲𝗺 𝗱𝗮𝘁𝗮, 𝗳𝗶𝗹𝗲𝘀, 𝗲𝘃𝗲𝗻 𝗱𝗲𝘁𝗮𝗶𝗹𝘀 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 𝗶𝗻𝘀𝘁𝗮𝗻𝗰𝗲 𝘁𝗵𝗲 𝗚𝗣𝗧 𝘄𝗮𝘀 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗼𝗻. The shocking part is that every new session spun up a fresh instance with new files — some from other users. That’s right. If people had uploaded docs to a public GPT, with the right prompt injection, I could read them. This made one thing super clear: 𝗱𝗼𝗻’𝘁 𝗲𝗻𝗮𝗯𝗹𝗲 𝗰𝗼𝗱𝗲 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 if you're making your GPT public. It’s a simple switch that could save your data from getting exposed. 🔐 ___ In my next post, I’ll break down exactly what kind of environment these GPTs are running in — from the Python version and system setup to just how much power the code interpreter really has. If you thought prompt injection was wild… wait until you see what’s under the hood. #AIPrompt #CyberSecurity #GPTs #PromptInjection #TechTips

Confused About Google’s Agent2Agent vs MCP? You’re Not Alone (Let's break it down) There’s been a lot of confusion about Google’s new A2A protocol. Will it replace MCP? Is it better? Should you switch? The truth? They serve entirely different purposes. Let’s break it down: 🚧 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗔𝟮𝗔: Agents communicate privately—peer-to-peer—via standard web protocols. Imagine it as a secure group chat. 𝗠𝗖𝗣: Classic client-server setup. Agents connect to tools via HTTP or even stdio. 💬 Communication Style 𝗔𝟮𝗔: Conversations between agents are flexible and collaborative. Like brainstorming in a war room. 𝗠𝗖𝗣: Everything’s schema-driven. Inputs and outputs follow strict rules—like blueprints in a factory. 🧠 𝗧𝗮𝘀𝗸 𝗛𝗮𝗻𝗱𝗹𝗶𝗻𝗴 𝗔𝟮𝗔: Is designed for long-term, complex workflows across multiple agents. 𝗠𝗖𝗣: It works best for simple, atomic tasks such as API calls and DB queries. 🕵️‍♂️ 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆 𝗔𝟮𝗔: Agents share “Agent Cards” to advertise capabilities. 𝗠𝗖𝗣: Uses JSON Schemas to define what each tool can do. 🔐 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗔𝟮𝗔: Enterprise-grade protection baked in (OAuth-level). Secure by design. 𝗠𝗖𝗣: You’ll need to bolt on security separately—especially for remote use. 🎯 𝗪𝗵𝗲𝗻 𝘁𝗼 𝗨𝘀𝗲 𝗪𝗵𝗮𝘁? → Use 𝗔𝟮𝗔 when you want agents to talk to each other. (horizontal scaling) → Use 𝗠𝗖𝗣 when you want a vertical scaling agent talking to multiple tools. You’ll need both if you're building real, scalable AI systems. Have you already explored A2A? What's been your experience with it so far? Credits for the amazing explanatory image to @rakeshgohel01 #AIagents #A2AvsMCP #AgentProtocol #ArtificialIntelligence #TechExplained