Your AI Agent Isn’t the Problem. Your Junk Drawer Is.

The bottleneck with agentic AI is almost never the model. It is the junk drawer you point it at.

The smartest agent on the market will still produce garbage if you point it at a mess. That is the part almost nobody wants to hear, because it is so much easier to blame the model.

I watch GTM leaders do this constantly. They try Claude Code, or OpenAI’s Codex, or Anthropic’s Cowork, get a sloppy first result, and conclude the tool isn’t ready. The model is fine. The model is, frankly, astonishing. The problem is they dropped a brilliant operator into a junk drawer of files with no labels and no system, then judged the work that came back out.

Here is the thesis, and I’ll defend it the whole way down: the bottleneck with agentic AI is almost never the intelligence. It is the operating environment you put the intelligence into.

The new hire nobody onboarded

Think about the best person you ever hired. Sharp, fast, eager, ready to go on day one. Now imagine you sat them at a desk, pointed at a wall of unlabeled boxes, and said “the files are in there somewhere,” gave them no idea where finished work goes, no sense of how your team writes things or names things, and then walked away.

A great hire would still try. They’d dig through the boxes, guess at your conventions, produce something. And it would be wrong in a dozen small ways, because you never told them what right looks like. You’d look at the output and think, “maybe this person isn’t as good as I thought.”

That is exactly what happens with agentic tools. The agent is the brilliant new hire. The repo or the folder is the desk and the boxes. And most people skip onboarding entirely, then blame the hire.

Here’s the twist. With a human, that onboarding is fuzzy and slow. It lives in hallway conversations and Slack threads and the slow osmosis of “how we do things here.” With an agent, the onboarding can be a file. One file. Written once, read at the start of every session, applied perfectly every time. The thing that takes a human three months to absorb, you can hand an agent in three hundred lines. That is a gift most people are leaving on the table.

Stop prompting. Start operating a system.

The first mistake is thinking the skill you’re building is prompting. It isn’t. Prompting is what you do in a chat window when you want a one-off answer. Working with an agentic tool is something different, and the leaders who get this pull away fast from the ones who don’t.

You are not prompting anymore. You are operating a system.

In that system, the repo or the working folder is the operating system. It is the environment the agent lives inside, the place where context, history, and output all accumulate. And the conventions file is the constitution. It is the document that tells the agent what the rules are, where things go, and what “done” looks like in your world.

Every serious tool now has a version of this constitution:

• Claude Code reads a CLAUDE.md file at the start of every session. It’s the standing brief the agent gets before it touches anything.

• OpenAI’s Codex uses AGENTS.md, which the team describes as a README for agents. Codex walks from the repo root downward and merges these files hierarchically, so a rule at the top applies everywhere and a rule deeper in applies locally.

• Anthropic’s Cowork, the desktop agentic workspace that operates inside folders you authorize, holds the same thing inside Projects: instructions, scheduled tasks, context, and memory.

Different names, same job. The tool changes. The principle does not. And this is the part that separates people who get leverage from people who get frustrated: a tool is something you buy, a system is something you build. The agent is the tool. The constitution and the folder structure are the system. If you only ever shop for tools, you’ll keep wondering why the magic doesn’t show up. The magic is in the system you wrap around the tool.

The one habit that compounds

If you do nothing else, do this. Keep a conventions file, and give everything a place to live. Then maintain both as the work grows.

That sounds almost too simple to matter. It is the most underrated move in the entire space, and the evidence backs it up. A study comparing human-written conventions files against ones the model generated for itself found the human-curated versions won. The machine-generated files actually reduced task success in five of eight settings tested. Read that again. Letting the agent write its own rulebook made it worse most of the time. The judgment about what matters, what’s non-obvious, what your team actually cares about, that still has to come from you. The agent executes the constitution beautifully. It is not yet the right author of it.

So write a lean one. The teams shipping with Codex say the same thing the Claude Code teams say: keep it tight, focus on the non-obvious rules, commit it and review it like code, because that’s what it is. For CLAUDE.md specifically, keep it under roughly two hundred lines, because the agent reads it every single session and bloat costs you. Boris Cherny, who built Claude Code, has a rule I’ve adopted wholesale: anytime the agent does something wrong, add a line to the file so it never does it again. That’s it. That’s the flywheel. Every mistake becomes a permanent correction instead of a recurring annoyance.

There’s a principle I live by in my own system, and it applies perfectly here: if it won’t exist next session, write it down now. The agent has no memory of yesterday unless you gave it one. The folder is its memory. The conventions file is its judgment. Every time you fix something verbally and don’t write it down, you are paying to teach the same lesson twice.

The compounding works like this. Day one, your conventions file is thin and your output is rough. You correct a few things, you write them down. Day thirty, the file knows where everything goes and how you like it, and the output lands clean on the first try. The system got smarter while you slept. That is the whole game, and almost nobody plays it on purpose.

Where to start