System 1 vs System 2 for AI Agents: Fast Calls for Routine Decisions, Slow Thinking for Hard Ones

A support agent makes hundreds of small decisions before lunch. Is this a billing question or a bug report? Is the refund inside the limit? Does this customer need a person? Is this email a request or a phishing attempt? Almost none of these take deep thought, and each one still needs an answer. When nothing the agent is trusted with can give one, it stops and asks a human.

That is where autonomy leaks away. Executives tend to worry about the hard calls, but the volume sits in the easy ones. If a person confirms one routine decision in ten, the agent is a fast assistant with a busy supervisor.

A new class of fast decision models has made the routine calls cheap to automate. The model is the cheap part. How much an agent does alone depends on a short list of choices that only your company can make.

Fast and slow, in a new hire and in a model

Daniel Kahneman's Thinking, Fast and Slow separates fast, automatic judgment (System 1) from slow, deliberate reasoning (System 2). A good new hire runs on both. Asked where an invoice goes, she answers in seconds. Asked whether to waive a clause in a customer contract, she stops, reads, and finds her manager. Swap the two and you get a hire who deliberates for ten minutes over every invoice, or one who waives clauses on instinct.

Agents have mostly had one speed. Every decision, trivial or not, went through a large language model that writes out its answer token by token, at a cost of seconds and money per call.

On September 15, 2026, TypeSafe AI released Jev in early access and called it a "System One model." Jev does not write text. It reads a piece of state, such as a support ticket, answers typed questions about it (pick an option, score on a scale, give the probability that a statement is true), and returns each answer with a confidence value. TypeSafe describes it as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

Five weeks, a crowded field

Other vendors shipped models of the same kind within weeks. Analyst Kai Waehner counts hyperscalers, AI labs, data platforms and more than a dozen open source projects. The market mostly settled on a different name than TypeSafe's: decision models. Five of the launches:

Every number in that table comes from the company selling the model. Each vendor also ranks first on its own benchmark, which is possible because each picks the subset of tests that suits its interface.

What independent testing shows so far

The independent evidence is thin and mixed. An October arXiv study tested Jev and an open-weight model, Laya, on 11 decision points taken from agent harnesses, such as choosing a tool or screening an input for prompt injection. Jev was significantly more accurate on nine of the eleven. Neither beat chance at zero-shot model routing.

The more useful part of the paper is the authors auditing their own work. A saving they first reported as 23.9% shrank to 4.3% once the cost of the pre-screening step was counted. A confidence threshold set to allow 5% misses allowed up to 17% on cases it had not seen.

Vendors add their own cautions. Microsoft's model card says Decision-1 is not designed to be the sole basis for decisions about credit, employment, housing, insurance, education, healthcare or legal rights. Waehner notes that calibration differs a lot between models and often needs a fit on your own data before the confidence values mean what they claim.

These models also answer from a fixed menu. They do not write, plan, or handle a situation nobody put on the menu. Open-ended reasoning belongs with a slower model, and the hardest calls belong with a person.

What it changes for an executive

If the model is a commodity, the leverage moves to what surrounds it. Four decisions are yours.

Name the routine calls. List what the role decides every day: route this, qualify that, approve anything under a limit. Most roles have fewer than a dozen. They are the calls a person currently double-checks.

Write the line for each one. A decision model returns a probability, and someone has to say what the company does with it. At 97% confidence, proceed. At 80%, take a second look. Below that, ask a person. This is the same shape as a delegated spending limit, and it belongs in writing for the same reason: the agent should not have to guess, and the manager should not have to reread.

Test the line on your own cases. The arXiv result is the warning. A threshold tuned on one sample misses more on the next. Run it on real past cases the agent has not seen before you widen what it may do alone.

Keep the record, and keep the model swappable. Log each decision with its question, its answer, its confidence and the model that produced it, so the lines can be tightened or loosened on evidence. Vendors are converging on one request format (Clef accepts Jev-style requests), so build for a model change to be a configuration change, not a rewrite.

Octavus applies the same logic to an agent's actions. Each Octavus Agent follows approval and escalation rules written in the company Handbook, so a person decides in advance which actions it may take alone, which it drafts for review, and where it must stop and ask. Our post on autonomous agent guardrails shows how to draw those lines, and the support guide on setting autonomy, approval, and escalation rules is the place to start.

The fast layer will be cheap and plentiful by the time most companies finish evaluating it. The slow part is deciding which calls are routine and how much confidence is enough, and that takes a meeting and a page of writing. The companies that have it written down first are the ones whose agents get allowed to run.

Sources

• TypeSafe AI: Introducing System One Models and Jev (Sep 15, 2026)
• Kai Waehner: How Decision Models Became a New AI Category After Jev (Oct 20, 2026)
• arXiv 2610.02267: Fast Models, Slow Evidence (Oct 1, 2026)
• Strands Agents: Introducing Strands Decider 2B
• OpenAI: Decisions API guide and public beta announcement
• Microsoft: Microsoft-Decision-1

Getting Started

Hire a pre-built Octavus Agent, give it the same tools as the role, and write down what it may do alone. For the basics, start with what autonomous AI agents are.