Fast Agents Produce Bad Results Faster

A lot of effort goes into improving agent output speed - streaming, caching, "thinking" indicators. Yet, none of it matters if the agent's output is not actually useful. This talk examines why capable agents consistently fail to deliver on their promise: and it's not because they lack capability, but because we've optimised for the wrong thing. The core problem is a speed paradox: agents are designed to optimise for speed to output rather than speed to understanding. Speed isn't what makes an agent valuable—useful, trustworthy results are. The talk is structured around three lessons learned the hard way building production agents at Flinn. Each lesson is grounded in a real system: the trust paradox was illustrated by our multi-channel medical device complaints handler, the failure of typical trust mechanisms (legibility, boundaries, reversibility) through our autonomous regulatory monitoring agent, and the question of whether chat is the right interface for agentic systems is explored through our clinical writing co-pilot. Across all three systems, the pattern was the same: we had to slow the agent down to make it useful. Participants leave with a framework for evaluating agents through the lens of trust architecture rather than technical capability, specific strategies for building understanding before execution, and concrete examples of what breaks in production and what works instead.