The best way to get an agent to use an SDK appropriately isn’t to give it documentation or an llms.txt file. Instead download the SDK add the SDK source directory to your agents workspace and tell the agent it has access to the code for the SDK but it can’t make changes.
Autonomous AI agents are no longer just tools; they're actors. This shift demands a new focus on security. In my new post, "Putting Up the Guardrails," I break down the threats and solutions, from prompt injection to layered defenses. #AI#LLMs#Security
Treat your AI agent like a smart intern: brilliant, but needs clear direction. Vague goals get vague results.
Part 5 of The Agentic Shift, I explore how to architect—not just prompt—an agent's behavior by separating its core mission, session context, and tools.
The Agentic Shift continues! 🚀 Part 4: "An Agent's Toolkit" is LIVE. Discover how AI agents use tools to transform from conversationalists into capable partners, executing multi-step workflows. Essential reading for anyone building with #AI! #LLMs#Developers 🔗 allen.hutchison.org/2025/10/04/a...
You’d never tell a junior dev, "Build me a login system," and walk away. You’d provide context, constraints, and examples. Why do we treat our AI partners any differently?
Coding with agents sometimes feels like an old schoo video game. There are puzzles to solve, traps to avoid, and sometimes when you get to the end you find out that the princess is in another castle.
Running two different agents simultaneously in the same branch feels a little bit like asking for trouble. They're working on different files through, so it'll be fine, right?
The ability for models to code in python is so much better than their ability to code in other languages. It's really shocking when you start a python project after working in typescript for awile.
I use some of my personal repos to test different CLI agents and their github integrations, but holy cow is it overwhelming to get feedback from four different bots when I submit a PR.
