Michael

What simple experiments or lessons from your research should they start applying right now — before agentic systems become mainstream? 

Start by running controlled experiments in simulated environments, such as Magnetic Marketplace, before deploying agents in real-world settings. Focus on stress-testing areas where current models struggle: handling choice overload, negotiating effectively, and collaborating toward shared goals. Use diagnostic tools such as the MCP Interviewer to identify tool interference and protocol issues early. Just as important, design with transparency and human oversight in mind, building layers that allow users to monitor and intervene. But most importantly, developing AI responsibly starts by asking questions about what can go wrong with the systems we are developing, how to test them for such risks, and how to mitigate them. By asking these questions throughout the lifecycle of the systems we develop, we can uncover rough edges and inefficiencies—and ensure that as agentic systems scale, they do so with reliability, fairness, and trust at their core.