From Vibe-Coded to Trusted: One Agent Through Build, Scale, Govern and Optimize on Gemini Enterprise
- Category
- AI & ML
- Format
- Session — 40 minutes
- Audience level
- Intermediate
Building an agent is easy now. Trusting it is the hard part. I take one vibe-coded agent, an ambient expense approver, through the four themes of Gemini Enterprise Agent Platform, using real results from my own project. Build: an ADK 2.0 graph where code makes the rules, Gemini only assesses risk, and a human approves anything risky; a prompt injection never reaches the model. Scale: one command to Agent Runtime, with sessions that survive a human pause. Govern: Agent Identity and an Agent Gateway with Model Armor set to fail closed, proven with requests that must be blocked (200, 403, 403). Optimize: an eval gate where a deterministic check caught a PII leak the LLM judges missed, and an online monitor where one prompt change lifted groundedness from 0.70 to 0.96 on the same 12 requests. Includes what went wrong and what is still not covered. Code and setup guides on GitHub. Key takeaways: 1. Let code decide what you know; use the model only where judgment helps, and keep a human on risky decisions. 2. Keep policy outside the model, fail closed, and test with a request that should be blocked. 3. Put a deterministic check beside your LLM judge: an average can hide 1 bad case.
Speakers
Jonathan Zhuo
GCP Principal Architect (Sr. Director) · Capgemini
Jonathan Zhuo is a Principal Cloud & AI Architect (Sr. Director) at Capgemini with 30+ years of industry experience and deep Google Cloud expertise — 7 GCP professional certifications, two publications on the Google Cloud website, and back-to-back Google Cloud Partner All-Star (2023–2024) and Champion Innovator (2024–2025) awards. Previously Chief Architect and Senior Google Cloud Practice Leader at Deloitte, Jonathan has guided Fortune 500 enterprises, federal agencies, and healthcare organizations through large-scale cloud and AI transformations. As an empty nester, he enjoys traveling, lifelong learning, and mentoring the next generation of architecture leaders.