An inbox sorting itself. A website linking itself. A board handing out its own cards.
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed
What you're watching: four of the ten builds back to back — the inbox, the website link map, the task board and the context meter — each one replaying real Jev decisions.
§2The twist
The model doing all of it can't write a single word
§3Who built it
TypeSafe AI — from Diogo Almeida, one of the people who co-invented ChatGPT
Tweet · the launch
“Why have superhuman chat models not led to AGI?”
This is Diogo Almeida's launch post. He says he spent the last two years in stealth building a new way to train models, and a new type of model came out of it: Jev.
§4The claim
20 to 200 times faster. 40 to 400 times cheaper. About a tenth of a second.
those three are TypeSafe's numbersmy runs from my desk: about 0.35 s per call, through OpenRouterOpenRouter lists it at $0.042 per million tokens in, output free
§5What it does, in one breath
You hand it a situation and a question with set answers. It picks one.
Tweet · OpenRouter
“A typed decision with a probability attached”
OpenRouter put Jev live in beta this week. Their description is the cleanest one I've seen: it takes your app's state plus a typed question, and it returns a typed decision with a probability. No prompting for JSON, nothing to parse.
§6System One
Red light, you stop. You don't write a paragraph about the red light first.
§7The problem
Every agent you run uses the paragraph brain for everything
Seconds go by. A full model call gets spent. And the answer was one word.
§8Why that changes what you can put on a screen
When a decision costs almost nothing, you can make thousands of them live
§9Three question types
Choice, Score and Noul — you'll see all three in the builds
§101 · Choice
Pick one from a list — with a probability for every option
§112 · Score
Rate something against levels you set yourself
§123 · Noul
Yes or no — with a probability from zero to one
§13All three at once
Ask all three about the same thing in one request. They come back together.
real Jev run · my machine · replayed at reading pace
What you're watching: one real email goes in with three questions — a Choice, a Score and a Noul — and all three answers come back from a single request.
my call: 3 answers · 0.98 s · $0.0000226
§14Right. Ten builds.
Each one is something you could sit and watch
§15Build 1 · the inbox that sorts itself
200 emails drop into folders — and a small pile lands in “needs you”
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · sample inbox, fictional senders▶ open the live demo
What you're watching: 200 emails fly into four folders one after another, and the ones Jev wasn't sure about land in a fifth pile for you.
my run: 200 emails · 5.24 s · $0.004123 landed in “needs you” — you check 23, not 200Riley Brown: 500 emails · seconds · 3.5 cents
Tweet · the proof
500 emails, three and a half cents
Riley Brown ran 500 emails through Jev. It classified them in seconds and the whole job cost 3.5 cents. So we already know this shape works, and what it costs.
THINKING IT?“What if it puts an important email in the wrong folder?”
That's what the confidence number is for. Anything Jev is sure about moves on its own, and anything it's unsure about goes in the pile for you.
On my first pass the pile was 37 emails. I rewrote the four folder descriptions so they didn't overlap, ran it again, and the pile dropped to 23.
The same three steps, every time
1
Describe the decision in plain English“Which single folder should this email go in?”
2
List the optionsReply · Needs research · Wait · Flag for me — each with one plain sentence saying what belongs there.
3
Set the confidence line0.4. Above it, the email moves on its own. Below it, the email waits in “needs you”.
§16Build 2 · a keyword list colouring itself in
2,000 rows fill with colour from the top down — and grey is the column you read
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · my real Search Console export▶ open the live demo
What you're watching: 2,000 real searches from my own Search Console get an intent colour one row at a time — blue, green, orange, and grey for the ones Jev isn't sure about.
my run: 2,000 real keywords · 30.8 s · $0.048951 informational · 274 commercial · 284 transactional · 491 greyHassan: 1,018 papers · 24 categories · $0.08 · 256 ms each
Tweet · the proof
1,018 documents, eight cents
Hassan sorted 1,018 research papers into 24 topics with Jev. Eight cents total, and about a quarter of a second per paper. Swap papers for keywords and the job is identical.
The same three steps, every time
1
Describe the decision in plain English“What is the search intent of this keyword?” and “Where should it live on the site?” — both in one call.
2
List the optionsInformational · Commercial · Transactional. Then: one of my closest existing pages, or “needs a new page”.
3
Set the confidence line0.5. Below it, the row goes grey — and grey is the only column you actually read.
§17Build 3 · a lead board with confidence scores
Leads sort into columns — and a handful turn red
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · sample leads, fictional people▶ open the live demo
What you're watching: 90 leads slide into strong, medium and weak, and the red column fills with leads whose outreach message doesn't match the person.
my run: 90 leads · 2.3 s · $0.001911 mismatches caught — good message, wrong personRomàn: 700 leads · 40 seconds · 9 cents
Tweet · the proof
700 leads, forty seconds, nine cents
Romàn gave Jev 700 high-intent leads with personalised messages. In 40 seconds it predicted how each message would land, gave each a confidence score, and flagged the lead-to-message mismatches.
The same three steps, every time
1
Describe the decision in plain English“How strong is this lead?” and “Does this outreach message actually match this person?”
2
List the optionsA score on three levels you write — weak, medium, strong — plus a yes or no on the match.
3
Set the confidence lineMatch under 0.5 goes to the red column. That red column is the outreach that was quietly wasting your time.
§18Build 4 · a whole website linking itself
Every page is a dot. Lines appear. And some dots stay unconnected.
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · 385 real pages of agentos.guide▶ open the live demo
What you're watching: all 385 pages of this website as dots, with a line drawn each time Jev places an internal link — and a pink ring on every page it chose to leave alone.
my run: 385 pages · 6.83 s · 295 links · 90 left alone · $0.011Borja: 586 pages · 45.1 s · 584 links · 139 refused · 21 centsClaude Opus 5, same clock: 21 pages
Tweet · the proof
586 pages in 45.1 seconds — and 139 it refused to link
Borja ran Jev over his whole site. It rebuilt the internal link map in 45.1 seconds, placed 584 links, and left 139 pages alone because nothing honestly fit. Claude Opus 5 on the same clock got through 21 pages.
Those unconnected dots are the proof the tool is thinking.
The same three steps, every time
1
Describe the decision in plain English“Which other page should this page link to, if any?”
2
List the optionsThe ten closest pages on the site — plus one more option: “none of these honestly fits”.
3
Set the confidence line“None” is the line. Being allowed to say nothing fits is what stops links going everywhere.
Every draft stops at a light before it goes anywhere
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · three Noul answers per request · sample drafts▶ open the live demo
What you're watching: 24 drafts queue through a traffic light; three probabilities come back for each one, and it goes to publish, to review, or back to the writer.
my run: 24 drafts · 72 answers · 1.1 s · $0.000612 green · 6 amber · 6 redit caught “500% ROI” and “100K subscribers in 6 months” as unsourced
The same three steps, every time
1
Describe the decision in plain English“Does this answer the search it was written for?” · “Does it make a claim with no source?” · “Are the internal links sensible?”
2
List the optionsYes or no, three times, in one call.
3
Set the confidence lineGreen: all three clear, publish and index. Amber: anything borderline waits for you. Red: a clear fail goes back to the writing model with a note.
§21Build 6 · the model router with a live cost counter
“Choose the cheapest model that can finish this job.”
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · spend bars use example prices▶ open the live demo
What you're watching: 40 requests slide to a small fast model or a big smart model, while two counters show what you spent against what you would have spent.
my run: 40 requests routed · 1.07 s · $0.000728 went to the small model · 12 to the big onethe dollar bars are an example: 10k tokens a request, $0.25 vs $15 per million
Article · LangChain
“Building a Harness with Jev”
LangChain shipped a ready-made piece for this. You describe in plain English what each model is good at, and Jev reads the request and picks. Sydney Runkle's post links the full write-up.
The same three steps, every time
1
Describe the decision in plain English“Choose the cheapest model that can finish this job.”
2
List the optionsEach model, described in plain English — small edits and lookups for the cheap one, hard decisions for the expensive one.
3
Set the confidence lineWhen Jev isn't sure, send the job to the big model. You only lose a little money, never the result.
§22Build 7 · the context meter that drops in one second
You run the plugin, and the bar falls off a cliff
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · sample 120-call agent session▶ open the live demo
What you're watching: a full context bar, one button press, a one-second scan of 120 tool calls, and the bar dropping from 950,000 tokens to 68,400.
my run: 120 tool calls scored · 2.9 s · $0.0019950,000 → 68,400 tokens at a 0.8 lineAlex Volkov: nearly 1M → 86K, in 1 second, inside Claude
Tweet · the idea
“Why is compaction still a summarization prompt?”
Tamara had the original idea: instant compaction. Jev scores every tool call in the history and drops what's irrelevant.
Tweet · the proof
Nearly one million tokens down to 86K
Alex Volkov ran it as a plugin inside Claude. It took one second, and his session went from nearly 1M tokens to 86K.
Tweet · the honest pushback
“Compaction isn't a filter”
Theo pushed back hard. His point: cleaning up history isn't the same as filtering it, and binning low scorers one by one can lose the trail that explains why the agent did what it did. He's right about the risk.
honest note: whether you delete or just reorder is still an open questionall of this is about a week old
The same three steps, every time
1
Describe the decision in plain English“Does this tool call still matter for the current task?”
2
List the optionsYes or no — asked once for every tool call sitting in the history.
3
Set the confidence line0.8 to keep. Set it high and you keep more of the trail. Set it low and the bar drops further.
§23Build 8 · the competitor monitor wall
A wall of dark tiles. Through the day, a few go bright.
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · sample change feed, fictional competitors▶ open the live demo
What you're watching: 16 competitor tiles stay dark while small changes get scored and filed, and a tile only lights up when a change scores above the line.
Describe the decision in plain English“Does this change matter to us?”
2
List the optionsYes or no — with one sentence describing what “matters” means for your business.
3
Set the confidence line0.75. Only tiles above your line light up. Everything else gets scored and quietly filed.
§24Build 9 · the browser you talk to
Jev decides. A small model types. The browser clicks.
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · a real headless browser on my own site▶ open the live demo
What you're watching: a request goes in, Jev picks one link out of 150 on a real page, the browser opens it, and Jev confirms the right page is open.
my run: 2 steps · 150 options · 1.4 s · $0.0003Browser Use: flights found in 7 s for $0.00391,092 → 101 browser commands · 25% fasterhonest note: it finds flights, it doesn't book themthe 7-second clock starts after the first page load
Tweet · the proof
“Browser Use + Jev = ultrafast”
Gregor Zunic from Browser Use posted this at 1x speed. After every click they rebuild the list of things you can click, Jev picks from that fresh list, and a small writing model fills in any box that needs typing.
Tweet · the full tutorial
The voice-controlled browser, built step by step
Moritz Kremb's tutorial builds the voice-controlled version at 5:59, plus an AI memory demo after it.
The same three steps, every time
1
Describe the decision in plain English“Which thing should be clicked next?” and “Is the page they asked for already open?”
2
List the optionsEverything clickable on the page right now. The list is rebuilt after every click.
3
Set the confidence line0.8 on “done”. Below it, the loop keeps going.
§25Build 10 · the task board that hands out its own cards
Twenty cards drop in and sort themselves. You didn't touch anything.
real Jev calls · typesafe/jev-1.13 through OpenRouter · replayed at viewing speed · sample cards▶ open the live demo
What you're watching: 20 cards drop onto the board and slide into the Claude Code, Hermes and OpenClaw lanes, the two cards Jev wasn't sure about slide to “you”, and the agents start working.
my run: 20 cards · 0.81 s · $0.0004“Decide refund for angry enterprise customer” → 0.59 → your lane18 cards assigned without you
The same three steps, every time
1
Describe the decision in plain English“Which available agent should take this card?”
2
List the optionsThe agents that are actually available right now — not a fixed list — each described in one line.
3
Set the confidence line0.75. Above it, the card moves. Below it, the card waits for you. That's what lets you leave the board running while you go for a walk.
§26What actually changed
Three things
§27One
Writing and deciding used to be the same job, at the same price
§28Two
Volume stopped being the limit
§29Three
Sure, it moves. Not sure, it asks you.
§30None of this needed a developer
Describe a decision. Write out the options. Set a line. That's the whole thing.
Wrong: “This is developer territory. It's not for a business like mine.”
Right: Every build on this page was a sorting, scoring or routing job described in plain English.
Wrong: “I can't trust an AI to make decisions for me.”
Right: You set the line. Above it, the thing moves. Below it, it comes to you. You only ever check the unsure pile.
Wrong: “I'll wait until it's out of beta.”
Right: The boring decision you already make a hundred times a week is safe to start on today — and you learn the skill while it's early.
Don't take my word for it
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries.
THINKING IT?“Doesn't running an agent operating system burn a fortune in tokens?”
No. The everyday work runs on free local models on your own machine, free APIs slot in for more, and the heavy work drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and the Agent OS plugs straight into it.
A deciding model like Jev pushes that even lower — all ten demos on this page cost seven cents — and the Boardroom has full token-efficiency tutorials.
§33Ten builds. Every one of them something you can watch.
Pick the one that hurts most right now and start there
Get the AI agent system and learn how to get more customers and leads.
The Agent OS zip, the decision-layer walkthroughs and the 30-day roadmap live inside the AI Profit Boardroom, with four coaching calls a week to set it up with you.
