Older post
Tools and Strategies for Agentic Development (Into the Hopper Podcast)
A conversation with Tim Hopper on spec-driven development, Beads for task tracking, and how AI agents have changed my planning habits.
January 12, 20262 min read
About 20% through chapter one of my new O'Reilly book on Agent Memory, and I'm constantly having to edit out asides I have. As many of you know, the act of writing encourages the connection of different ideas together. But that doesn't stop me from posting about them here!
One thing I keep thinking about is the language we use for our tools. To call LLM runtimes "agents" and frame their context window as "memory", we're doing something closer to what Judith Butler called performativity than mere description. Our repeated framings don't just describe the technology we're building, but are actively helping constitute it. What we call these systems shapes what we build, what the user expects from them, and what we plan to build on top of them.
Which gives me pause. Not because our language is too strong. But because it is too weak for what we're building. Human memory is intangible and ungraspable. Agentic memory is different. We can inspect it, shape it, version it, compose it. That makes what engineers are building something more than memory, not less. Our language hasn't caught up to the technology. Part of what I want this book to do is give readers the practical tools to build these systems, and the conceptual tools to think past the words we currently have for them.
More asides like this will show up here as I write. The ones that don't make it into the book end up there. I send them out at econoben.dev/book#subscribe.