AI in the Open: Agentic AI
AI in the Open is a webinar series exploring how open source and artificial intelligence are shaping the future of technology. Hear from industry leaders, practitioners, and subject matter experts on emerging trends, real-world use cases, and the opportunities and challenges organizations face as AI evolves.
Each session offers practical insights you can apply to your own work, while connecting broader industry perspectives across open source and AI. Join us live every third Thursday of the month, or watch sessions on demand on the Red Hat Open YouTube channel.
The next edition takes place on Thursday, October 29.
Talk #1
Title: Defining Optimality and Efficiency in the age of Agentic Authorship
Speakers: Joseph Groenenboom and Craig Magina
In this age of agentic AI, having it writing the kernels for you could be a big time saver. But, how can you tell if an Agentic AI generated kernel is performing to the best of its ability when you don’t have a comparable hand written kernel to compare against? Benchmarking can tell us how fast it is, but without anything to compare it against, you are left wondering if it’s making optimal use of the hardware.
In this talk, we will show you how to determine the optimal hardware utilization for a given kernel, generated by Agentic AI. Our session will demonstrate a robust methodology for determining if an Agentic AI can generate a reasonably performant kernel and show how close the performance comes to optimal. We can find the answer by profiling the kernel to see how the kernel utilizes the hardware.
Attendees will learn how to look at profiling metrics to gain deeper insights into performance bottlenecks to improve a generated kernel. Finally, we will analyze the results from different kernels to see some of the problem spaces where Agentic AI excels in kernel generation (well-covered kernel spaces) versus those where it struggles due to lack of training data (new kernel spaces).
Talk #2
Title: From Token Efficiency to Trustworthy AI Coding Agents
Speaker: Cédric Legendre
AI coding agents can generate code faster than ever, but how much of that activity translates into meaningful engineering progress?
This talk introduces TER (Token Efficiency Ratio), an open-source project that evolved from measuring token efficiency into analyzing AI coding sessions through Lean principles, identifying wasted effort, and producing evidence-based recommendations.
Building on TER, we explore three related experimental projects: TGV, for intelligent task routing between local and cloud models; GARE, for autonomous execution, repair, and recovery; and BALISE, for independent benchmarking and validation.
Together, these projects illustrate a broader engineering philosophy: measure, decide, execute, and verify; shifting the focus from generating more code to delivering reliable, measurable outcomes.