People say JEV is the cheap, fast, accurate decision maker. The best at it.
A session is one ongoing chat with one language model. That chat makes hundreds of decisions. So the idea sounds obvious: keep the same chat, keep the same model, and send every decision to JEV. JEV answers. The model keeps working. The chat never starts over, so the memory stays.
I thought that was the smart setup. It is not.
Inside the roomOne sessionJEV points
>
Best at pointing at a door is not the same as best person to ask while you are already inside the
room.
Picture three doors
A hallway has three doors. The signs say Easy, Normal, and Hard.
JEV stands in the hallway and points at one door. That is the whole job. Point. Fast. Done.
You walk in. Inside the room, one helper works with you. The helper hears everything you say in that room. That room is the session.
Here is the bad move. You are already in the room. A choice comes up. You walk back to the hallway. You explain the whole room to JEV. You wait. JEV points. You walk back and tell your helper. Your helper still does the work.
Do that for every choice and you spend the day in the hallway.
The hallway tax //
Each trip costs words, waiting, and an answer that stays in the chat. After enough trips, the chat
is full of "JEV said Normal," and your helper has to read all of that.
Why JEV inside a session is a bad idea
JEV only points. It picks from a menu you already wrote: A, B, or C. It does not do the work. Your helper still does the work after the point. You added a trip. You did not remove one.
JEV is not in the room. It only knows what you repeat in the hallway. Repeat too little, and the point is a guess. Repeat the whole chat every time, and you pay for the same story twice.
Cheap plus a lot of trips is expensive. Each trip costs words, waiting, and an answer that stays in the chat. After enough trips, the chat is full of "JEV said Normal," and your helper has to read all of that. The sticker price was small. The session got bigger and slower.
JEV cannot swap the helper. If it points at a different door, you have to leave the room. The new room is empty. The helper does not come with you. You tell the story again, and you lose whatever you forget. That loss is worse than staying with the helper you already have.
So the logic breaks in one place. Best at pointing at a door is not the same as best person to ask while you are already inside the room.
When it is a good idea
Let JEV point once, before you walk in. Then stay in the room until the work is done.
Or skip the room. Some jobs are only a label: which folder, which team, which button. JEV can do that job alone, with a short note and a menu. Those jobs are in Ten jobs where JEV alone is the right tool.
Point once //
No session yet: let JEV pick the door. Then stay with that one model until the work is done.
What I do now
No session yet: I let JEV pick the door once. Then I stay with that one model.
The whole job is a label: I ask JEV once, I take the answer, I move on. No chat.
A session is already open: I do not ask JEV. The helper in the room already has the story.
>
A session is already open: I do not ask JEV. The helper in the room already has the story.
Stay in the room //
JEV inside a session adds a trip. The helper still does the work. Point once, before you walk in,
or hand JEV a label and leave the hallway.
Latest Blog Posts
Field Note2026-09-26
Ten jobs where JEV alone is the right tool
JEV fits when the decision is the whole job. A short note, a menu, one point. Ten jobs with that shape, and the test that keeps a live session out of it.
Manifesto2026-09-19
The Agent Gateway: Stop Asking Your AI to Be a Psychic
Your AI is not a mind reader. The Agent Gateway is a front door: one knock, then the rooms, the locks, and a stamp that proves it actually showed up. Not more pages. A door.
Investigation2026-09-15
The NemoClaw Librarien Next Door
NemoClaw MCP memory that works today: Sparky answers on port 8888. Catalog cards, not copies. A workaround, and three small asks for NVIDIA.
Investigation2026-09-13
What is NemoClaw doing with semantic search?
NemoClaw 0.0.123 on an NVIDIA Spark: qwen3-embedding sits idle, the sandbox still searches by exact words, and the docs describe a path the image does not open. The problem is real. I need to investigate further.
A
B
C
This article is part of a Semantic Cluster. All links are managed by the Digital Architect AI.