# System 1 and System 2: thinking fast and slow with Jev and language models

How Jev's intuition guides a reasoning model through complex agentic tasks

Canonical: https://agentrun.ai/blog/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models
Published: 2026-10-27T18:21:26.000Z
Author: Miguel Rios Berrios

_How Jev's intuition guides a reasoning model through complex agentic tasks_

Daniel Kahneman's _Thinking, Fast and Slow_ (2011) splits how we think in two, and both halves are smart. System 1 is the expert's glance: a chess master sees a strong move, a doctor has a hunch before the tests come back. Kahneman quotes Herbert Simon on it: "Intuition is nothing more and nothing less than recognition." System 2 is the one that sits down and reasons it through, like doing your taxes. Most of the time it's System 1 that tells System 2 where to put its attention.

We build agents the same way. Their System 1 is Jev, TypeSafe's model for quick typed judgments (yes or no, a score, or a box). It can look at a frame and tell in about a quarter of a second whether what you're after is there. System 2 is a smart language model that reasons through the harder parts. In this post, System 2 is Gemini 3.8 Flash because we're working with video and audio, but it could be Claude Sonnet, Opus or Fable, or GPT Terra, Sol or Astra. Because intuition is fast, it's also cheap, a few thousandths of a cent per look. Reasoning over video is the slowest and most expensive call an agent makes because you pay for every second you send. So Jev is the North Star. It tells System 2 where to look, then checks what System 2 found

[Explore System 1 and System 2 agent loops](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/visualizations/agent-loops.html)

We put the two to work on two jobs: a needle in a haystack, and the kind of tedious knowledge work analysts do every day.

## Finding a needle in a \[video\] haystack

We wanted the visual equivalent of a classic needle-in-a-haystack test: bury a brief event inside a long video, then see how vision models - and this System 1/System 2 loop - behave when they have to find the exact moment. The “needle” was a golden robot.

The film is _The Middleton Family at the New York World's Fair_, a 1939 Westinghouse picture that's now in the public domain. About 34 minutes in, Elektro, Westinghouse's seven-foot golden robot, walks out onto a balcony above a crowd, small enough to miss. Before any model saw the film, we wrote down when he shows up and drew a box around him. A model got it right if it named the same moment, give or take a second, and put its box where ours was.

We tried three ways: give the whole film to the video model (A), ask an LLM about every frame (B), or let Jev's intuition say where to look first (C). In C, Jev glances at 40 frames, narrows in on the stretch where the robot starts to look likely, and hands Gemini the one minute worth reasoning about.

[Explore How the loop found the robot](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/visualizations/robot-search.html)

## Results

-   **A. Video model alone** - Found: 3 of 3; When: a second late; Cost: $0.24; Time: 31 s
-   **B. One call per frame** - Found: no; When: a brass gauge, 11 minutes in; Cost: $4.13; Time: 10 min 39 s
-   **C. Jev, then Gemini** - Found: 3 of 3; When: the exact second; Cost: $0.009; Time: 24 s

![Video frames compare a video model alone, one call per frame and Jev-guided Gemini when searching for the hidden robot.](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/6970c820618527e9.jpg)

I expected A to miss, because we'd found video models to be pretty bad at exact moments. It got him in all three runs, a second late, because it waits until he's fully lit.

The difference is the price. C got the exact second for under a cent, 27 times cheaper than A and a little faster, because A reasoned over all 55 minutes and C only reasoned over the minute Jev pointed at. On a library of 10,000 hours that's about $2,600 against about $95.

B is the version most people write first, and it was wrong. A model that sees one frame at a time with no context says yes to anything golden: 13 times before the robot showed up, and 2 more after he left. Its first yes was a brass gauge, 11 minutes in. It also burned $4.13 and four and a half hours of model time. Jev looked at that same gauge and gave it 0.34, and the loop only sends System 2 to frames at 0.7 or more.

[Explore Jev frame scores](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/visualizations/frame-scores.html)

## Every payment method a merchant takes, with proof

Ask a compliance analyst which payment methods a merchant accepts, and you've just taken 15 to 30 minutes of their day. They walk the site, add something to the cart, open every dropdown and popover, get to checkout, and keep a list of every card brand, wallet and pay-later option they find, with a screenshot for each one. Card brands hide behind a "+5", wallets hide behind "Show more options", and by the third merchant it's easy to lose track. It's the kind of job where System 1 and System 2 are better together, and this is how we do it.

[Explore Payment-method detection pipeline](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/visualizations/payment-pipeline.html)

An agent does the walking: a pi agent with pi-browser on a Browserbase browser. On [allbirds.com](http://allbirds.com/) it took a screenshot after every page, scroll and opened control, 37 in under six minutes, for $0.86. It never types anything, never signs in and never presses pay. pi-browser's own recorder won't keep frames while payment fields it can't read are on screen, so the agent's screenshots are the recording.

Then Jev takes the lead. It glances at every frame and says which screens are worth reasoning about, and the split is clean: 0.01 to 0.03 on screens without a payment method, 0.99 or 1.00 on screens with one. Gemini reasons over only those screens and names every method and the frame where it's clearest. Jev goes back over each one, points at where it sits on the frame and confirms it's there (all 14 scored between 0.77 and 1.00), and Gemini takes the close look and draws the box. Anything that doesn't check out goes back around.

What comes out is an evidence pack: every method, the frame it's on and a box around it. On [allbirds.com](http://allbirds.com/) that's 14 methods, the same 14 we counted by hand, in all three runs, for about 10 cents and about two minutes on top of the browse.

![Contact sheet of payment-provider logos and screen examples, including card networks and digital wallets.](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/a919f5e21ef8b382.jpg)

For a compliance team the check is the part that builds trust, so we fed Jev a batch of claims where we knew some were wrong. 38 claimed methods came from a model that had read the recording in one go, and 13 of them pointed at the home page because it had read seconds as minutes. Jev scored all 13 at 0.01 or 0.02 and flagged every one for another look. The 76 checks cost under a cent.

[Explore Jev payment-method claim scores](/blog/imported-writing/system-1-and-system-2-thinking-fast-and-slow-with-jev-and-language-models/visualizations/claim-scores.html)

## Where we're taking it

Jev's intuition shouldn't wait for the recording. It can ride along with the browsing agent and say which control is hiding more methods (that "+5", the "Show more options" link). When a logo is a few pixels wide, it can point Gemini at the patch of screen to zoom into, or at the neighbouring frame where the logo is bigger. It's the same split, earlier in the loop. Run that over a whole portfolio of merchants and the analyst gets a folder of evidence packs instead of an afternoon of clicking.

## System 1 and System 2 in AgentRun

We use this split all over AgentRun, and vision is only one place. In pi-vision, vision.judge is Jev, vision.see is the video model, and the find skill teaches the agent which to reach for first. In AgentRun's workflow language a judge, pick, sift or route node is one Jev call, and an agent node is a model with tools. Jev routes a run to the right prompt before the first turn, or sifts a thousand items down to the ten worth reading, and the reasoning model does the reading.

Both examples follow the same four beats:

1.  System 1 takes a quick look and says where to look.
2.  System 2 reasons there.
3.  System 1 uses that result and checks it.
4.  System 2 gives the end result, one we can stand behind.

Jev knows at a glance which minute of a film has the robot in it, or which screen has a payment logo on it. Gemini works out the exact second and reads every logo. Put them together and you get agents that do the job better than either one could alone, for a lot less. On the robot that meant the exact second for under a cent, and on the checkout an evidence pack an analyst can hand to an auditor.