Sathiesh Veera is an AI Solutions Architect with more than 17 years of software engineering experience. Across his roles at various Fortune 500 organizations he has designed and delivered large-scale systems from unified IAM platforms and data lakehouse platforms built on open standards, to dynamic decisioning and rules engines for financial transactions, and AI data layers and RAG pipelines for LLM-powered applications. An IEEE Senior Member and AWS Community Builder, Sathiesh is a peer-reviewed researcher focused on identity and data protection and data security, with recent work on securing enterprise data for LLM applicationsl. He serves on AI Governance Board at his current role and mentors early-career engineers. Across every role, the throughline of his work has been the same lesson: the systems that last are built by people who respect the engineering fundamentals and assume ahead of time that things will go wrong  and it's that hard-won perspective, more than any single technology, that he enjoys sharing with the community.  

Title: The Agent Did What? Building Trust and Control into Autonomous AI Systems - Abstract:  The first time I watched an agent take an action nobody explicitly asked it to take, I understood that autonomy changes the whole safety conversation. A model that answers a question is one thing; a system that decides, calls tools, and acts on your behalf is another — and the failure modes are far less forgiving. This talk is about giving autonomous AI systems room to be useful without giving them room to do harm. I'll walk through where agentic systems go wrong — over-broad tool access, chained actions with no checkpoint, decisions made on bad or manipulated inputs — and the controls that keep them accountable: scoped permissions, human-in-the-loop gates on irreversible actions, action logging you can actually audit, and blast-radius limits for when things go sideways. The lesson I keep coming back to is that trust in an autonomous system isn't something you hope for; it's something you engineer, one boundary at a time. You'll leave able to let agents do real work while staying firmly in control of what they're allowed to touch.