AI agents with a human touch are way better than general agentic systems that try to automate everything. The human touch happens at phases like prompt design, context engineering, agent architecture, and, more importantly, evaluation. Domain expertise matters so much.

I’m excited to finally launch this. My MCP server now offers real-time stock market news. Available tools: • get news • get stock prices • get key metrics • get financial statements Took about a month to get the MCP server right, but it's working well now. More financial tools coming soon.

MCP is getting so good!

What is context engineering❓ And why is everyone talking about it...👇 Context engineering is rapidly becoming a crucial skill for AI engineers. It's no longer just about clever prompting; it's about the systematic orchestration of context. 🔷 The Problem: Most AI agents fail not because the models are bad, but because they lack the right context to succeed. Think about it: LLMs aren't mind readers. They can only work with what you give them. Context engineering involves creating dynamic systems that offer: - The right information - The right tools - In the right format This ensures the LLM can effectively complete the task. 🔶 Why Traditional Prompt Engineering not enough: Early on, we focused on "magic words" to coax better responses. But as AI applications grow complex, complete and structured context matters far more than clever phrasing. 🔷 4 Key Components of a Context Engineering System: 1️⃣ Dynamic Information Flow Context comes from multiple sources: users, previous interactions, external data, tool calls. Your system needs to pull it all together intelligently. 2️⃣ Smart Tool Access If your AI needs external information or actions, give it the right tools. Format the outputs so they're maximally digestible. 3️⃣ Memory Management - Short-term: Summarize long conversations - Long-term: Remember user preferences across sessions 4️⃣ Format Optimization A short, descriptive error message beats a massive JSON blob every time. 🔷 The Bottom Line Context engineering is becoming the new core skill because it addresses the real bottleneck: not model capability, but information architecture. As models get better, context quality becomes the limiting factor. I'll share more as things evolve and become more concrete! Stay tuned!! 🙌 ____ If you found it insightful, reshare with your network. Find me → @akshay_pachaar ✔️ For more insights and tutorials on LLMs, AI Agents, and Machine Learning!

if you give me the code, i can execute which means the value in execution is protected by code but AI is making coding easier so specifications become the moat because if one has specs, they can produce the code and they can execute

Agents casually completing captchas is low key hilarious

imagine a modular conversational interface where you: - select model of choice - pick your search provider - choose MCP tools - access your file systems - manage your memory, with security and privacy - pay as you go for any services used

If anyone’s as confused about software licenses as I was, here is how I see them now. Software licenses control what people can do with your code. MIT license lets people do anything. Say you build a JSON parsing library with MIT license. A startup can take your code, modify it, and sell it as part of their paid API service without giving you anything back. They just need to keep your copyright notice in the code comments. React uses MIT, so Facebook could have made it proprietary, but they chose to keep it open. This is the "do whatever you want" license. GPL forces sharing. Imagine you create a web framework under GPL. If someone modifies your framework to add new features, they must release those modifications as GPL too. They can't sell a proprietary version. WordPress uses GPL, so all WordPress themes and plugins must be GPL too. If you build a commercial CMS on top of WordPress, your entire CMS becomes GPL. This keeps everything open source. Apache 2.0 handles patents better. You write a machine learning library and someone contributes an algorithm. Later, they try to sue users for patent infringement on that algorithm. With Apache license, they automatically lose their rights to use your library. It's protection against patent trolls. Kubernetes uses Apache 2.0 because cloud companies worry about patent issues. BSD is basically MIT with different words. You create a networking library under BSD. Same rules as MIT, people can do whatever they want. FreeBSD uses this, which is why Apple could take BSD code and put it in macOS without releasing macOS source code. Most developers just use MIT now because it's clearer. LGPL is GPL for libraries. You build a PDF generation library under LGPL. Companies can use your library in their proprietary software without making their whole app LGPL. But if they modify your library code itself, those changes must be LGPL. VLC media player uses LGPL so other apps can include video playback without becoming GPL. Proprietary means you control everything. You build a database engine and keep it proprietary. Companies pay you license fees to use it. They can't see the source code, can't modify it, can't redistribute it. Oracle Database works this way. You make money from licensing but limit who can use it. Dual licensing gives options. You release a database under both GPL and commercial licenses. Open source projects use the GPL version for free. Companies that don't want GPL restrictions pay for the commercial license. MySQL does this. Startups use free MySQL, big companies often buy commercial licenses. Public domain gives up all rights. You create a hash function and put it in public domain. Anyone can do anything with it, no restrictions, no attribution required. SQLite does this. Government agencies love it because there's zero legal risk. Building a library you want everyone to use? Pick MIT. Want to ensure improvements come back to the community? Use GPL. Worried about patents in enterprise software? Go with Apache. Building a business around your code? Stay proprietary. Want maximum freedom for users? Try public domain. The key is matching your goals with license restrictions. Don't just copy what other projects do without understanding why they chose that license.

MCP is like the HTTP for agents, and feels like the foundation of a new internet. We shipped the blog post that answers the most common questions i've been getting from developers, partners and enterprises. Shoutout to MCP's creator @dsp_ for collaborating on this!

this is huge from google research. highly tech paper but much needed i guess