|
Anirban Lahiri (Arndit Ltd., Cambridge, United Kingdom)
 Agentic AI: A Friend or Foe
Agentic AI — systems capable of autonomously working towards goals, making decisions, and taking actions in dynamic environments—are rapidly transitioning from research prototypes to real-world actors. This talk examines a central question: are these systems best understood as powerful collaborators that amplify human capability, or as emerging risks that steal our jobs, challenge our ability to maintain control and accountability? This talk explores the conditions under which agentic AI becomes either friend or foe.
We begin by outlining some technical foundations of agentic systems, including planning loops, tool use, memory, and multi-agent coordination. These capabilities enable applications ranging from scientific discovery and software engineering to autonomous operations in finance, healthcare, and infrastructure. Case studies highlight tangible benefits such as accelerated research cycles, improved decision support, and adaptive systems that respond to complex, changing environments.
However, increased autonomy introduces new classes of risk. The talk examines failure modes such as misaligned objectives, reward hacking, emergent coordination between agents, and brittle behavior under distributional shift. It also addresses broader societal concerns, including labor displacement, concentration of power, opacity in decision-making, and challenges in assigning responsibility when systems act independently. Particular attention is given to the gap between system capability and our current tools for evaluation, monitoring, and governance.
The central argument is that agentic AI is neither inherently beneficial nor inherently dangerous; its trajectory depends on deliberate choices in design, deployment, and oversight. The talk outlines some pathways for responsible agentic AI, emphasizing interpretability, controllability, human-in-the-loop design, and robust governance mechanisms. By aligning technical innovation with societal values, we can shape agentic AI systems that act as reliable partners rather than unpredictable adversaries.
Anirban Lahiri is a Data Solutions Architect at Kainos and Founding Director of Arndit Ltd., Cambridge, United Kingdom. Anirban has been developing advanced technology in the Computing Industry over more than 2 decades spanning 7 countries across 3 continents (US, Europe and Asia) . He has worked for many multinationals and is credited for taking numerous ideas from the concept to successful product launch for companies including Philips-NXP Research, ARM, Siemens, Xaar, Kaleao. He is credited with bringing to market the Big-Little Architecture commonly found in mobile phones and tablets nowadays and the first steps to building an ARM based supercomputer. He has authored more than 10 patents for his inventions and a number of books/book-chapters as well as 30+ articles published in numerous conferences and journals. He is also closely associated with teaching and research at Cambridge University. He holds an MBA from Imperial College London and MS degree from Indian Institute of Technology, Kharagpur. He has also been a Visiting Fellow, at University of Texas, Austin and Stanford University.
|
