When someone asks what tools I use to run operations across companies without being a coder, they expect a list. I used to give them one. I would name the AI models, the project management app, the deploy platform, the document system, and watch their eyes glaze at the length of it. The list was accurate and useless. It described what was installed on my machine and told you nothing about how work actually moves through my day.
The real stack is not the tools. It is the roles. Four jobs need doing in any operation I run, and each one needs a different kind of tool. Once I stopped collecting software and started assigning roles, the constant swapping stopped and the work started shipping.
The four roles
Direction is the first role, and it is mine. Someone has to decide what "done" means, write the task clearly enough for a machine to act on it, and check the result against a real standard. The tools I direct are three AI coding agents: Claude, Codex, and GLM. Each has a different strength, and routing work to the right one is not an optimization trick. It is the difference between a stack that is sustainable and one that burns money on the wrong problems. Claude handles the judgment work, drafting and review. Codex is fast on well-defined tasks. GLM, the cheapest of the three, handles routine work where I do not need to pay for reasoning I am going to verify myself anyway.
Memory is the second role. Work has to live somewhere that outlasts the conversation it was discussed in. For agent coordination, that place is a shared Notion database I call the Agent Queue. Every task is a ticket with its own context, claim rules, and a written receipt when it is done. For leads, it is a Google Sheet. For accounting records, it is a self-hosted document system that files everything automatically. The principle is the same across all three: the record does not depend on anyone remembering it.
Shipping is the third role. Something has to put the work in front of real people. Our company website deploys to Vercel the moment code lands on the main branch. I do not run a deploy script. I push, and a minute later the change is live. That tight loop is what makes directing AI practical. If every change required a developer's calendar and a release window, the speed advantage of AI would die in the waiting.
Alerting is the fourth role. When something happens that a human needs to know about, the system has to interrupt the right person in the place they already are. For our team, that is Slack. For me, running agents through a messaging gateway, that is Telegram. The tool does not matter as long as the alert lands where someone will actually see it instead of inside a dashboard nobody opens.
The trap I fell into
For the first few months I did it backwards. I collected tools. Every time a new AI model launched or a friend recommended an app, I added it to the stack and expected it to fix the fragmentation I was feeling. It did the opposite. Each new tool solved a local problem and created a new integration problem, because nothing connected to anything. I had four places where tasks could live and no single source of truth. I had three AI agents and no way to stop two of them from working the same task at the same time.
The stack did not start working because I found better tools. It started working when I defined the roles and made each tool own exactly one. The Notion queue became the only place work lives between agents. The Google Sheet became the only place leads live. Vercel became the only path to production. Once each role had one owner, the tools stopped competing and the connections between them became the actual system.
What this produces
I run operations across multiple companies. I ship production software I never wrote a line of. Our accounting records file themselves. Leads get answered before I wake up. None of that required me to become a coder. It required me to stop thinking about tools and start thinking about architecture.
What you can copy this week
If you are a founder building your own operation around AI tools, do not start with a tool list. Start with four questions. Who directs the work and checks it? Where does the record live so nobody has to remember? What puts the finished work in front of real people? Where does the alert go when something needs a human?
Answer those four, pick one tool per answer, and refuse to add a fifth until one of the four is clearly failing. The founders who ship with AI are not the ones with the longest tool lists. They are the ones with the cleanest architecture and the discipline to stop swapping.