Fast Decisions for AI Agents: Hands-On with Jev, a System 1 Decision Model

Most decisions inside an AI agent are not reasoning problems. Which team should handle this ticket? Is this message a prompt injection? Is this request safe to automate? How urgent is it? These are classifications, yet most teams send them to a full large language model. They pay in latency and tokens, then write fragile code to parse free-text answers back into decisions.

In this 40-minute hands-on workshop, we explore a different approach: a fast "System 1" decision layer that sits in front of your reasoning model. Working live in Google Colab, we will use Jev by TypeSafe AI to ask typed questions and get calibrated probabilities back in a single pass, with no text to parse.

What we will cover:
• Why every agent decision does not need an LLM, and where System 1 fits in Harness Engineering
• Jev's three question types: Noul (yes/no probability), Choice (categorical routing) and Score (ordinal ratings)
• Asking multiple questions in one parallel pass, and how it compares against a standard LLM classifier
• Building a ticket-triage agent that blocks prompt injections, automates safe actions, escalates urgent issues and routes the rest
• Tuning confidence thresholds live, and knowing when this approach is the wrong tool

Bring your challenge: In the final segment, submit your own messages to the triage agent, including attempts to trick the guardrail. If you manage to break it, even better: that is where the learning happens.

Who should attend: AI engineers, developers and data professionals building agents or LLM-powered applications. Basic Python familiarity is helpful; no prior experience with Jev is needed.

You will leave with: the complete Colab notebook to keep and run yourself, plus a design pattern you can apply to your own agents with any tool.

Date: Monday, 5 October 2026 Time: 7:00 PM–8:00 PM GST Format: Live online workshop Host: Mohammad Arshad Ahmed, Founder & CEO, Decoding Data Science