System one models bring reliable automation to ai engineering
Diogo Almeida and engineers from Glean, UiPath, and Workday demonstrate how fast, non-generative models like Jev enable reliable software automation at the AI Engineer Code Summit.
この記事は英語版のみ利用可能です。
At the AI Engineer Code Summit on November 12, 2026, industry leaders gathered to discuss a shift away from purely generative large language models toward faster, more deterministic alternatives. Diogo Almeida, co-founder of TypeSafe AI and lead creator of Jev, headlined the event with a call to integrate these "system one" models directly into application code for greater reliability.
What happened
The central theme of the summit was the rise of system one models, which differ significantly from the reasoning-heavy large language models that have dominated recent AI development. These newer models return structured choices, scores, or probabilities rather than generating open-ended text. They are notably faster and cheaper to run, addressing a critical pain point for engineers who need consistent performance in production environments. Almeida described this approach as "reliable intelligence," arguing that it should be used to automate virtually every aspect of software interaction where predictability is required.
Almeida emphasized that code is the primary language of automation and that system one models must be embedded within it. He criticized the current state of LLM-driven software, noting that while these models can perform like Einstein ninety-five percent of the time, they occasionally fail in ways reminiscent of Mr. Bean. This inconsistency makes full automation dangerous if developers cannot distinguish between high-quality and poor outputs. His keynote urged engineers to liberally sprinkle system one model calls throughout their codebases to create smarter, more robust applications without relying on fragile string concatenation techniques.
Following Almeida’s keynote, Chinmay Goyal from Glean presented practical applications of these models in enterprise search. Goyal explained that Glean uses system one models alongside traditional classifiers and LLMs to handle decision spaces that change frequently. The key advantage he highlighted was zero-shot flexibility, which allows the system to adapt to new taxonomies without requiring repeated fine-tuning. Goyal noted that end users do not need to know which model is powering a specific feature; they only care that the product works reliably. For Glean, hiding the complexity of model selection behind a seamless user experience remains the primary goal.
How it works
System one models operate by evaluating inputs against a fixed set of outcomes or scoring criteria rather than generating novel content token by token. This architectural difference allows them to execute thousands of calls in under a second, making them suitable for real-time decision-making tasks. At the summit, Chibi Vikram and Kevin Bernstein from UiPath demonstrated this capability through a document ranking tournament. Their system used Jev to select the most relevant chunks of text from small groups, advancing winners to subsequent rounds until the best matches were identified. This multi-step strategy leveraged the speed of the model to achieve high accuracy without the computational cost of a full generative pass.
Scott Roberts, a Principal Engineer at Workday, showcased another implementation where the model predicts user intent before an action is completed. In his prototype, the interface begins filling out a time-off card while the user is still typing. This shifts the software’s role from merely predicting what a user wants to say to anticipating what they want to do. By integrating fast model calls directly into the user interface logic, the application becomes proactive rather than reactive. These examples illustrate that system one models function as adjustable building blocks that gain intelligence through strategic composition rather than raw parameter count.
Key details
- System one models return structured data such as scores, probabilities, or discrete choices instead of open-ended text generation.
- Jev, released in September 2026 by TypeSafe AI, is cited as a primary example of this technology, offering speed and reliability advantages over reasoning LLMs.
- Glean utilizes these models for zero-shot flexibility, allowing their search taxonomy to evolve without constant model retraining.
- UiPath demonstrated a tournament-style ranking system where Jev processed thousands of calls in under one second to identify relevant document chunks.
- Workday prototyped an interface that predicts user actions in real-time, filling forms as users type to reduce friction.
- Diogo Almeida compared the current experimentation phase with system one models to the early days of the internet, encouraging developers to explore new software boundaries.
Why it matters
For software engineers and technical founders, the shift toward system one models represents a move from experimental prototypes to production-grade reliability. The unpredictability of generative models has long been a barrier to automating critical business processes. By using models that provide structured outputs, teams can build systems that behave like standard logic gates, ensuring that automation flows do not break due to hallucinations or erratic behavior. This reliability allows companies to automate more complex workflows with confidence, reducing the need for extensive human oversight.
Furthermore, the cost and latency benefits of these models make them viable for high-volume tasks that were previously too expensive to handle with large language models. Engineers can now embed intelligent decision-making directly into the core logic of their applications without worrying about prohibitive API costs or slow response times. This democratizes access to advanced AI capabilities, enabling smaller teams to build sophisticated features that were once the domain of well-funded research labs. The ability to iterate quickly on these models also means that software can adapt to changing business requirements without lengthy retraining cycles.
What you can do
- Identify parts of your application where structured decisions are needed, such as classification, ranking, or routing, and replace generative calls with system one models.
- Experiment with embedding model calls directly into your application code to handle real-time user intent prediction, similar to the Workday prototype.
- Use tournament-style ranking strategies to improve accuracy when processing large datasets, leveraging the speed of system one models to filter results efficiently.
- Evaluate your current AI workflows for opportunities to reduce latency and cost by switching from reasoning LLMs to faster, non-generative alternatives for simple tasks.
- Explore zero-shot classification capabilities to handle evolving taxonomies without the overhead of continuous model fine-tuning.
- Join the community of developers experimenting with these new frontiers, treating the current landscape as an opportunity to define best practices for reliable AI automation.



