Jev and the new decision level for AI agents – Unite.AI

Because System One models could separate fast judgment from slow reasoning

Many AI agents use a language model for almost all of their decisions. The language model chooses a tool, evaluates the results, determines whether it needs to move forward, and finally generates responses. Flexible; however, this process can be costly when yes or no decisions are repeated on a large scale. Unite.AI has previously discussed how agent workflows augment model calls, context, and retries. Each additional decision can add time and money before providing users with useful information.

Jev suggests dividing the task differently. Use a model built for limited judgments in which the set of responses is defined. Use a generative model for open reasoning and language. Jev suggests that the main thought here is not that all agents need to buy a new product. The key concept is that an agent does not need the same type of intelligence at all times.

What Jev really does

TypeSafe launched Jev in September 2026the first of their new System One models. Jev doesn’t write prose. Instead, you send it a status (such as a support message and user data). You can also send one or more questions with predefined answer types. Then Jev responds with typed answers and probabilities.

According to the official company documentationThere are three primitives for making judgments:

• Choice allows you to choose from predefined options.
• Point it allows you to evaluate something against a neat rubric.
• The new estimates the probability that a statement is true.

You can ask multiple independent questions about the same status within a single request.

For example, let’s say you’re handling a customer service issue. A system might want to figure out which team should handle this case. It can also determine how quickly someone needs to respond and see if the customer has requested a refund.

A chat model could potentially do all three tasks. However, it will need to provide the results to your app as a structured response. In contrast, Jev provides only those limited decisions. Your app will then decide what action to take next based on those decisions.

The architectural change matters more than the model

Most of these debates compare large models to small ones. Jev proposes an alternative border. Some steps involve language generation. Others are narrow judgments that the software can consume.

This creates a decision layer in the agent. The model will estimate. The software will enforce the policy. If the estimated probability exceeds a tested threshold and the action is low risk and reversible, the workflow can continue. If there is uncertainty in the results or if the action could have serious implications, the system may require human supervision. A reasoning model can help investigate uncertainty, but it does not replace the necessary human approval.

There are similarities to template routing, but there is one key difference. RouteLLM makes decisions about which of the two linguistic models to choose. Select between a stronger and weaker model to balance quality and price. A System One model produces limited judgments that code can use directly. These judgments can support model routing as well as other decisions within an agent.

Why agent loops are a natural solution

The nature of agent circuits makes them particularly suited to making numerous judgments at very small levels. These judgments help achieve the final result. In other words, agents have to make many “small” judgments after a user submits their question or request. These judgments occur before the response or output is returned.

An example would be deciding which tools to use, classifying retrieved records, and assessing risk. The system also determines whether there is sufficient evidence and whether the trial should continue. This will most likely happen repeatedly. Additionally, the delays between each cycle may increase over time.

This role for agent loops is exemplified by LangChain Jev integrationwhere Jev can perform both model routing and tool call controls. While Jev integrates at the edges of the generative model, the generative model itself continues to plan and generate content. This represents a much more realistic use case for Jev. It complements a general-purpose language model rather than replacing it.

Additionally, parallelizing questions also changes the way teams think about task decomposition. Specifically, teams are able to break down an ambiguous statement into multiple discrete evaluation questions. This can potentially result in a much shorter sequence of model calls. It can create a much easier workflow to evaluate. It also allows developers to use explicit business logic to combine the resulting judgments.

Generic language models can produce structured results and in some cases may be the best choice. For example, you may need to provide a determination and an explanation together. Therefore, Jev must demonstrate more than simple compliance with the scheme to be considered effective.

The effectiveness of Jev depends on achieving reductions in overall system latency. It also depends on producing useful probability estimates and stability of performance across different inputs. If Jev fails to deliver these benefits, choosing another model will only add further development and operational costs.

Does typed mean correct?

The language used when making statements about Jev also needs to be phrased carefully. Since the output space is defined in advance, the model should not return a made-up field or an unparsable paragraph. This eliminates one form of failure; it does not eliminate the semantic error. There is nothing to stop a system from naming the wrong department, assigning the wrong risk level, or declaring too much certainty. It can do all this while being completely type-agnostic.

Own by TypeSafe System One documentation makes an important distinction. Calibration is measured on groups of forecasts; does not guarantee the correctness of an individual forecast. In manufacturing, this has implications. Teams need to check whether the predicted probabilities match the results observed on their data.

Performance evidence remains early

TypeSafe reports response times of between 70 and 500 milliseconds. It also references substantial cost savings and speed improvements in internal workflow assessments. Furthermore, TypeSafe indicates that such headline earnings are likely close to the high end of real-world earnings. TypeSafe’s Workflow tests available to the public uses reference probabilities provided by other frontier models rather than truth labels. The results are useful for formulating hypotheses. The results cannot replace an independent test against a real workload.

A practical test before adoption

When building your first AI decision workflow, don’t choose the most critical decisions (for example, medical approvals or account suspensions). Instead, choose something that is very common, reversible, and easy for other people on the team to review. This includes, but is certainly not limited to ticket routing, document categorization, template selection, and low-risk quality assurance.

Four questions will help you judge whether it will work:

• Does the output have a finite number of possible answers?
• Can you clearly articulate the criteria for judging?
• Are there measurable results? Track the prediction, its probability, action and subsequent results. Check calibration regularly by comparing predicted probabilities with observed results.
• Do you have an alternative plan in case the automated decision making fails? Identify a specific point where you can use a reasoning model, ask for more information, or involve a human.

Your analysis should include the entire workflow, including decision making. Use metrics such as decision accuracy, abstain or escalation rates, total end-to-end processing time, cost per successfully completed task, and error impact. Test under adverse conditions: variable use of words, omission of relevant data, infrequent categories, and contradictory input. An optimized classifier that generates additional downstream costs is not an optimization.

The long-term lesson here

Whether Jev succeeds, changes significantly, or is replaced quickly, one thing remains constant. The architectural question remains. Is it necessary for every decision made by the machine to be rendered as generated language?

In many cases, the answer is “no”. In a manufacturing environment, a system using generative models can generate interpretations, plans, and explanations. Using limited decision models, the same system can route, score, and gate. Code can continue to dictate acceptable threshold values ​​and permissions. Human beings should remain responsible for decisions that affect the lives of others.

While this represents a less dramatic prospect than an autonomous model that performs all tasks reliably, it reflects how reliable systems are created. The next advance in agent performance may depend on selecting those areas of the system where thinking takes longer. Other areas need quick decisions, while others require no action.