Files Are All You Need

Files Are All You Need is one of those ideas that sounds almost too simple… until you’ve lived through the pain it’s trying to solve.

If you’ve spent any real time working with AI agents, especially coding or research agents, you’ve probably felt that creeping anxiety when context starts to disappear. You know the moment. The agent forgets something important, rewrites work you already did, or confidently goes in the wrong direction. I’ve been there. It’s frustrating in a very human way.

In the original article by Jerry Liu from LlamaIndex, which you can read here https://www.llamaindex.ai/blog/files-are-all-you-need, the core idea is refreshingly grounded. Instead of stuffing agents with endless tools or massive context windows, we treat files as the main interface for memory and understanding.

Think about it like this. When you work on a complex project, you don’t keep everything in your head. You write notes. You create documents. You organize folders. Agents can work the same way.

Modern coding agents like Claude Code and Cursor already do this. They store conversations, plans, and research in files. When context gets tight, the agent doesn’t panic. It looks things up, reads a file, and keeps going. Almost like a coworker flipping through shared docs.

The article highlights three emerging patterns, but two really stand out in day to day use. First, files act as long term memory. Instead of losing context, agents write it down and revisit it later. Second, files are replacing clumsy retrieval systems. Rather than pulling random chunks of data, agents search, scan, and read files intentionally, the way you or I would.

What’s exciting is where this leads. Fewer tools. Simpler agents. Clearer workflows. Research.md, plan.md, implementation files. Messy, practical, human.

It feels like a return to basics, and honestly, that’s a good thing. As these systems mature, agents that can calmly read, write, and reason over files might end up being the most reliable collaborators we’ve had yet.