Your playbook for making AI real, practical, and valuable.
Jie Tao, DSc, an associate professor of business analytics and the director of the AI and Technology Institute at Fairfield University’s Charles F. Dolan School of Business, provides listeners with a practical guide to mastering AI for real results.
Each episode delivers actionable tools, proven frameworks, and real-world case studies to help leaders and innovators leverage AI for business growth and career success. Explore topics from the Practical AI Playbook to insights from Fairfield Dolan's AI and Tech Institute and discover powerful AI applications shaping the future.
Episodes
7 days ago
7 days ago
14 min
Few people build their own AI agent harness. Most rely on Claude Code or OpenCode. Dr. Jie Tao, Director of Fairfield Dolan's AI & Tech Institute, built one anyway, and Dr. Philip Maymin asks what he learned. The answer is three open questions: how much to delegate, how to manage knowledge, and whether today's harnesses have hit their ceiling.
First, the numbers. Tao tried 15 ways to regulate his agents, from plan-first and "grill me" to a knowledge base and deterministic logic. He mapped 12 failure modes and added four safeguards from his engineering background, including a kill switch with a hard-rule trip line. He has solved six. The other six, drift above all, keep tripping, and AI can't fix them either.
Tao wants to delegate. But delegation without control is pointless, and control depends on getting knowledge out of his head and into the agent, which, after 60-plus years of trying, we still can't do well. Until we can, he treats the harness as a colleague, not an autopilot.
On the ceiling: nearly all our control happens at authoring time, with almost nothing governing an agent while you use it.
"Teaching a student, the loss is acceptable if you're putting enough effort. Teaching an agent, the loss sometimes is not acceptable."