Definition
The half that decides without writing
System One is the half of AI that decides without writing. You hand a model something to look at, plus the list of answers it is allowed to give, and it picks one off that list and says how sure it is. No sentence is produced at any point, which is the whole of the difference.
The pieces have their own names. A model that works this way is a System One model, and TypeSafe released the first public one, Jev, in September 2026. The loop that keeps asking it, step after step, until a job is finished is a System One harness. The two running together against a real system are a System One agent.
The name is borrowed from psychology. System 1 there is the thinking that happens on its own: you recognise a face, read a room, or know an answer is wrong before you can say why. System 2 is the thinking you have to do on purpose, effortful and one step at a time and slow enough that you notice you are doing it. Psychology writes the digits; the software names are written out in words, and this page follows that.
System Two, in the software sense, is what nearly every model you have used is doing: working through the problem in words, one token after another, and handing back prose, whether or not it shows that working. Only one side of the split is a name anyone sells under, though. System One is a category with products in it; System Two stays a borrowed contrast rather than a label on anything for sale.
Where the names come from
Kahneman named two characters, not two parts of a brain
The distinction is older than the labels. Psychologists Keith Stanovich and Richard West introduced System 1 and System 2 as names in 2000, and Daniel Kahneman adopted them in Thinking, Fast and Slow in 2011, crediting them in the book, which is where most people met them.
He presented the two systems as characters in a story rather than as parts of a brain, and said plainly that they are fictitious: there is no System 1 organ, and he was not claiming the head is divided the way the labels suggest. Dual-process researchers have argued ever since about whether the properties grouped under each label really belong together. The labels survive that argument because the contrast they draw is useful, not because the boundary was ever found.
What survived is the contrast: fast against slow, effortless against effortful, many things at once against one step at a time, wordless against verbal. Software now has two kinds of model that fall on either side of every one of those lines.
In software
AI has used this split twice, in opposite directions
The first borrowing was a research agenda. At NeurIPS in 2019 Yoshua Bengio gave an invited talk titled From System 1 Deep Learning to System 2 Deep Learning, arguing that deep learning had done well on the System 1 tasks, the perceptual and intuitive ones people perform without noticing, and that the next step was System 2: reasoning, planning, causality, generalisation that holds up outside the training distribution.
What arrived afterwards took the vocabulary rather than the programme. Prompting a model to work step by step, and then training models to spend tokens deliberating before they answer, made deliberation something you can buy by the token. That is not the causal, systematically generalising System 2 Bengio set out, and the label stuck to it anyway, which is why the term now means two different things depending on who is using it.
The second borrowing runs the other way, and it raises the obvious objection: if a neural network is fast and intuitive, why does a language model count as the slow one? Because the speed is in the network and the deliberation is in the interface. The network runs the same fast parallel pass whatever you ask it. What is serial is the stream of tokens coming out, one at a time, and that is what earns these models the label. Fast judgment was never missing from the network. What was missing was a way to get it out without the stream.
Ask a language model for a judgment and you get a sentence, and your code has to turn that sentence back into a value: a string matched against a list, a number parsed out of prose, a retry when the wording drifts. A System One model returns the value itself, instead of something your code has to recover from prose.
The probability that comes with it is what lets your code decide whether to act at all. Classifiers have returned scores for decades; the difference is that the answer set is written per request rather than trained in, so the number applies to a question nobody trained a model to answer.
The differences
System One and System Two, compared
The two differ in what you send, what comes back, and how the answer is produced.
| System One | System Two | |
|---|---|---|
| What you send | A state to look at, plus questions whose permitted answers are enumerated in advance | A prompt in natural language |
| What comes back | A typed value for each question. A choice or a score carries the probability of every permitted answer; a yes-or-no question carries the single probability that the answer is yes | Text, which may contain the answer |
| How it is produced | Every question in the request is evaluated at once, independently of the others | One token after another, each conditioned on the ones before it |
| How sure it is | A probability you can read and compare against a threshold. A choice or a score also returns a confidence value, which says how concentrated the whole distribution is and is not the same number as the top answer's probability | Token log probabilities where the provider exposes them, which measure how likely each word was rather than how likely the answer is to be right, or a separate scoring pass |
| Staying inside a fixed answer set | Guaranteed by construction: it can only hand back an answer you listed | Achievable with constrained decoding or schema enforcement applied on top of generation |
| Typical latency | Tens to hundreds of milliseconds. TypeSafe publishes 70 to 500 ms for Jev, measured from its own machines near the service | Typically seconds, and longer again when the model deliberates before answering |
| What it cannot do | Write, explain, summarise, code, or say why it answered as it did | Reach the answer without first producing the tokens that contain it, which is where most of the latency and the cost sit |
| Where it fits | Frequent bounded decisions: route this, classify this, gate this, choose the next action | Anything whose output is language, and anything whose options cannot be listed in advance |
The latency figures are the vendor's own, measured under conditions it discloses, and are the kind of thing you can check yourself in an afternoon. The guarantee in the fixed-answer-set row is a guarantee about shape, not about truth: a System One model cannot return an answer you did not offer, and it can still pick the wrong one from those you did.
Together
The handover is the design decision
The two are a division of labour, not a choice: the constant, bounded, individually cheap decisions go to the fast model, and the ones that need language out, or that fall outside the set you could enumerate, go to the slow one.
What makes it work is that the handover happens at a named point. In a loop built on a System One model, asking for help is one of the actions on the menu, and a step whose weakest judgment falls below the bar set for that action runs nothing and is recorded with the whole distribution behind it. That record is the handover: a larger model, or a person, picks the run up knowing exactly where it stopped and how the belief was split when it did.
Kahneman's own picture of the two systems has roughly this shape. System 2 runs at low effort most of the time, monitoring, and is mobilised when System 1 meets something it cannot settle or produces something that needs overriding. The difference in software is that you get to choose the call site, write the bar down as a number, and read back later how often it fired.
Naming
Five things called System One
The name now sits on a concept, a model category, a loop, an agent and a specific piece of software, and they are not interchangeable.
| Name | What it refers to | Whose term it is |
|---|---|---|
| System 1 | The fast, automatic mode of thought. A concept from psychology, written with a digit, and not a claim about anything in software | Stanovich and West, 2000, popularised by Kahneman in 2011 |
| System One model | A model that answers typed questions with values and probabilities and generates no text. TypeSafe calls its Jev the first public model in the category | TypeSafe, introduced with Jev |
| System One harness | The loop around such a model: it holds the state, offers the actions that are possible right now, gates on the probability that comes back, and stops for a reason it can name | HarnessRouter |
| System One agent | A System One model and a System One harness together, pointed at an environment and given a goal. Neither half is an agent on its own | HarnessRouter |
| System One Harness | HarnessRouter's open-source implementation of that loop, released under Apache 2.0 | HarnessRouter |
Each of the last four has a page of its own: what is a System One model for the interface and how to read its numbers, what is a System One harness for what the loop has to do, what is a System One agent for the two working together, and System One Harness is open source for the release and the runs measured behind it.
Next
The decisions this fits
The shape fits decisions your software already makes hundreds of times a day and would rather not wake a person for. Which queue does this ticket belong in. Is this order ready to ship. Did this page change in a way that matters. Which tool should run next. Is this step finished. Today those are usually brittle rules that break on the first odd case, or a language model writing a word that your code immediately matches back against a fixed list.
The test is short. If you can write down every answer your code is prepared to accept, this shape fits. If you cannot, it does not.
One decision on its own is not much use. Something has to keep asking, step after step, until a job is finished, and a model doing that inside a loop built for it is a System One agent.
FAQ
Common questions
Is System One the same as System 1 in Kahneman?
It is named after Kahneman's System 1 and is not the same thing. That System 1 is a way of describing how people make fast judgments, and he was explicit that the two systems are fictitious characters rather than parts of a brain. A System One model is a product category: a model that returns typed answers with probabilities instead of text. The borrowed name points at a resemblance in behaviour, not at a shared mechanism, and nothing about the model was derived from the psychology.
Is System One thinking better than System Two thinking?
Neither is better; they answer different questions. System One is the right shape when you can write down every answer your code is prepared to accept and you need the answer often and quickly. System Two is the right shape when the output is language, or when the options cannot be enumerated in advance. Systems that use both are common, and the design question worth arguing about is where the handover sits rather than which half is smarter.
Who invented the term System One for AI?
Three different people, for three different things. The psychology terms are Stanovich and West's, from 2000, popularised by Kahneman in 2011. The framing was applied to machine learning most prominently in Yoshua Bengio's NeurIPS 2019 invited talk, From System 1 Deep Learning to System 2 Deep Learning, though he was not the first to reach for it. System One as the name of a model category is TypeSafe's, introduced with Jev, and it is a vendor category name one release old rather than an established model class with a published architecture. System One harness and System One agent are HarnessRouter's: the model category arrived with no name for the software that has to be built around it, and those two are defined in what is a System One harness and what is a System One agent.
Is a System One model just a faster language model?
No. A language model generates text one token at a time and the answer has to be recovered from what it wrote. A System One model evaluates every question in a request at once and returns the value itself, with the probability behind it, and produces no text. TypeSafe describes a different architecture and training objective, although it has not published details that would let anyone verify the internals.
Can a System One model and a System Two model work together?
That is what the probability is for. The fast model takes the frequent bounded decisions, and a step whose weakest judgment falls below its bar executes nothing and is recorded with the full distribution behind it. Repeated refusals stop the run with a named reason, at which point a larger model or a person picks it up. The handover works because the stopping point is explicit rather than inferred.
Should I write System 1 or System One?
Use the digits for the psychology, System 1 and System 2, which is how Kahneman and the research literature write it. Use the words for the software category, as in a System One model or a System One agent, which is how TypeSafe introduced it.
Put a harness behind your product
Run complete agent harnesses through one contract for tasks, sessions, streaming and artifacts, and keep the harness a request parameter rather than an architectural commitment.
Start building free