AI Explained: Why Do Smarter AI Agents Still Break?
Training a capable AI model is one problem. Getting an agent built on top of it to work is a different one. The two are more connected than they seem: the failures an agent runs into once it's deployed reveal exactly what to train into the next version of the model, delivering a better agent.
Juhi Parekh is GM of Key Frontier AGI Accounts at Turing, with firsthand experience working with frontier AI labs on post-training pipelines, benchmarks, and RL environments, and seeing how the agents built on those models hold up once enterprises deploy them. Join her for a conversation about what it takes to build an AI agent that works, and why even a capable one can still fail once it's live.
What you'll learn:
- How teams design test problems hard enough that a model fails, improving it before deployment
- Better reasoning, more reliable tool use, and staying coherent through long, multi-step tasks are letting agents take on far more complex work than they used to
- Why agents that pass benchmarks still fail at specific things once deployed, like generating documents or interpreting images correctly
AI Explained is our AMA series featuring experts on the most pressing issues facing Agentic and AI teams.