See it work
Sixteen small simulations of the same idea: a fast, cheap AI makes the small judgements, your code makes the decisions, and the expensive AI is only called when it is genuinely needed. The last six put it between AI agents, in six multi-agent architectures. Every page links to the Python it is built from.
The article · 24 min read Jev and System One models: a working engineer's guide to typed, calibrated AI decisions The thinking behind every simulation below — start here if you want the why.Customer conversations
3 simulationsMessages arriving from people who want something, and the question of who — or what — should deal with each one.
Support triage cascade
A thousand customer messages arrive. Most of them never need an expensive AI model — but something has to work out which ones do.
The shared inbox
One mailbox, three people half-watching it, and a purchase enquiry buried under four recruiters. Sorted before anyone opens it.
The second pair of eyes
A chatbot wrote a reply to a customer. Did it invent a number? Promise something? Actually answer? Check before it sends.
Sales and orders
2 simulationsThe front of the business: an enquiry that might be worth a quarter’s revenue, and an order that arrived as six unpunctuated words.
Scoring while they wait
Someone submits a form. Rate them on four things and route them before the page finishes loading. Drag the sliders to change what matters.
Orders by WhatsApp
For a lot of businesses the ordering system is a phone number. Jev judges what the message is; a regex reads the number. Neither does the other’s job.
Money and paperwork
2 simulationsThe back-office grind — the work that is boring, constant, and expensive to get wrong. This is where a calibrated confidence number earns its keep.
The bill that just arrived
Invoice or statement? Paid already? Does it match something we actually ordered? Your open orders go into the same call as the document.
A month of transactions
Eight hundred bank lines, and the only number that matters is whether the model knows which ones it is unsure about. Watch the accuracy move with the floor.
Around the models you already run
3 simulationsPatterns that sit in front of, behind, or between frontier models — screening what they read, checking what they do, and deciding which one gets the job.
The gatekeeper
Search returned ten documents. One is irrelevant, one is trying to hijack the answer. Screen them before the expensive model reads a word.
The checkpoint
An AI agent keeps asking to run commands. Check every one for four separate hazards, in less time than the command itself takes.
Picking the right brain
You pay for several AI models of different speeds and prices. Send each request to the cheapest one that can actually do the job.
Agentic systems
6 architectures · 6 simulationsJev between the agents
LLM agents do the work. Every decision between them is a small, typed question: pass this on, who goes next, hand this over, is it done, is it looping. Jev answers each in a quarter of a second, and your code holds the thresholds. One use case for each of the common architectures.
Start here · the overview Agentic systems: where Jev sits in all six architectures One diagram per architecture, what decides each step today, and where to be careful.In each small diagram the pink diamonds are Jev, blue boxes are LLM agents, green is code or done, and orange is a person.
Three desks, one claim
Intake, assessor and letter agents in a line. A gate at every handoff passes the work on, sends it back once with a reason, or stops for a person.
Four specialists, one verdict
Jev picks which of four specialist agents this supplier needs. They run at the same time, and a second call merges their findings into one decision.
The foreman
Jev is the supervisor. Every step it picks the next worker, checks the last output, and notices when the team is going round in circles.
Passing the customer on
On every customer message, stay with the current agent or hand over. A margin in code stops agents passing the customer back and forth.
Write, judge, rewrite
A writer agent drafts and Jev scores the draft on a five-part rubric. "Done" is a line of code, feedback names the failing checks, and the loop knows when to stop.
The agent on call
An incident agent proposes commands. Jev checks each one for hazards, and also asks whether the agent is repeating itself and whether the incident has already cleared.
The shape they all share
1. Your code gathers exactly the information the question needs. 2. Jev answers every question about it at once, and says how sure it is. 3. Your code decides what happens — the thresholds are numbers you can read, test and revert. 4. An expensive model is called only for the part that genuinely needs one.
Each simulation matches the runnable Python in
cookbooks/, linked from the top of every page. The numbers shown are illustrative,
not measured.