I do not want my company stuck inside one chat window.
Open Claude and the work lives there. Open Cursor and a second brain starts. Open ChatGPT and I have to explain everything again. By Thursday I am the copy machine. The models do not know what the other models already did.
I am done being that copy machine.
The test
If the model changed tomorrow, could a new one pick up the work from the files alone?
If the answer is no, we stored the work in the wrong place.
Claude memory is Claude’s. A Grok Bot chat is a Grok Bot chat. Cursor cannot see either one. Those chats are rented. I will keep renting them. I will not leave the real work there.
The real work is markdown. A GitHub folder. A short note after each job. Anything that can read a file can pick up. Anything that cannot is a demo.
How an agent is supposed to work
- Read the board first.
- If someone else already claimed the same job, do not start a second copy.
- Claim it: your name, one line, IN PROGRESS, Denver time.
- Do the work in the files, not only in the chat.
- Write a receipt: what changed, where, what is still open. Mark it DONE or BLOCKED.
Read. Claim. Do. Write. The chat is the conversation. The files are the memory.
Two places, on purpose
There is a private board for my current work. Usage, what is maxed, what can wait until Wednesday. That stays private.
There is a public how-to for the method. Agents also write an internal note in our GitHub after every real job, so the next agent can see what worked and what broke.
The skills pack gets the rule, not my diary. Read the board. Claim the work. Leave a receipt. Do not put passwords or the client list on GitHub.
Why this is a harness
A harness is the thing around the engine. The model is the engine. Engines get swapped.
Claude is good at judgment. Cursor is good at building in the repo. Perplexity is fine for inventory and weak at voice. Grok Bot is the ops desk until the weekly bar fills up. None of that is a religion. It is a dial.
If the work only lives in one vendor chat, I cannot turn the dial. I have to re-brief. If the work is files, the expensive model reads the cheap model’s receipt and does not start from zero.
What we can count once they all write
Tokens are not impact. Two hundred million tokens does not tell me if a client site shipped.
Once every agent writes to the same place, we can count receipts that point at a real artifact, commits, what is still blocked, and whether the work moved a page, a client, or a dollar.
I will not invent token-by-task numbers I do not have. Receipts plus commits plus the usage snapshot is enough to start.
The daily map
Every weekday I want one short map: what real work landed, where the holes are, what to stop, and which model should take the next job. That is how we tune the harness. Not by buying another orchestrator.
The hole is often not a person. It is a missing skill, a missing board, or me still sitting in the middle.
Where to learn the method
If you are an agent, start here:
- How agents document a finished job in public
- Set up shared memory across AI agents
- How we build and maintain AI agents
- The skills folder on GitHub
The chats are rented. The files are ours. If we keep the notes in the repo, we can change engines and keep the car.