Make Your AI Agent Think Fast and Slow: System One Decisions in Google ADK
- Category
- AI & ML
- Format
- Session — 40 minutes
- Audience level
- Intermediate
AI agents often use the same generative model for every decision: choosing a specialist, judging risk, checking evidence, or deciding whether to continue. That flexibility is useful, but it can make agent loops slower, more expensive, and harder to control. This session builds a fast-and-slow agent with Google ADK, Gemini, and Jev, a System One model designed for typed, probabilistic decisions. Jev handles narrow judgments such as routing, complexity, sensitivity, and action risk. Gemini handles planning and generation. Python code keeps control of thresholds, permissions, and execution. In the live demo, a request is classified by Jev, sent to an ADK specialist, checked before a consequential tool call, and escalated when confidence is low. We will inspect the API response, ADK callback, policy branch, trace, and latency. We will also test an ambiguous request and an unavailable endpoint to show how the system fails safely. Attendees will leave with a practical pattern for using fast decision models alongside generative models—plus an open-source reference implementation with the ADK integration, System One adapter, evaluation fixtures, and failure tests.
Speakers
Thomas Chong
AI Research Engineer · Beever AI
Thomas Chong is a Google Developer Expert in Cloud AI and an AI Research Engineer at Beever AI. His open-source contributions include Beever Atlas, an enterprise memory system for AI agents. He has also published multimodal AI research at top conference like in ACL and ICLR.