AI Distribution, Chips and Agent Frameworks: A March 2026 Livestream

In this March 20, 2026 livestream, Matthew Berman and Nick Wince discuss AI news and interview Jayden Clark, Microsoft’s R. Andrew Wall, Eliza Labs founder Shaw Walters and Sentient co-founder Himanshu Tyagi. The guests describe their own organizations and offer predictions, not independent product evaluations.

Clark and the hosts discuss distribution and attention as software becomes easier to build. Clark argues that attracting attention is more sustainable when a useful product establishes credibility first.

They discuss Nvidia’s GTC announcements and the relationship between agent usage and demand for compute. They note that Nvidia has a commercial incentive to encourage more compute consumption.

The hosts consider a claim about spending on AI tokens per engineer. Clark says the right amount depends on whether the tools help produce useful results rather than on a universal spending target.

The group discusses how productivity improvements might affect hiring. Their examples and forecasts are opinions; they do not establish how employment will change across the economy.

Microsoft’s R. Andrew Wall introduces the Azure Maia chip program. He describes designing hardware and software together and argues that offering both internal accelerators and third-party GPUs gives customers more choice.

Wall discusses coordination between infrastructure and model teams. The hosts ask about the economics of frontier inference, where a new capability can initially be expensive while older capabilities become cheaper over time.

Eliza Labs founder Shaw Walters describes an open agent framework and consumer-oriented applications built around it. He presents the products as a way to explore how agents interact in applications and games.

Walters discusses simulated scenarios and a proposed benchmark for scam resistance. He distinguishes performance in a particular game from the harder question of generalizing resistance to real-world scams.

The conversation turns to whether websites and payment systems should adapt to agents or whether better models will learn to use existing interfaces. Walters offers ideas for agent-to-agent communication and emphasizes security.

Sentient co-founder Himanshu Tyagi introduces Arena, an effort to compare and improve open agent harnesses. He says the aim includes generating open feedback and data about how systems attempt tasks.

Tyagi says model capability and the surrounding harness both matter. He discusses generating multiple candidate improvements, evaluating them and retaining useful ones, while noting that the first cohort had not yet produced final results.

Asked whether agents should discover existing skills or create their own, Tyagi says the answer depends on the amount and quality of feedback and evaluation data. He contrasts adapting an existing skill with generating one from evaluation feedback.

The show closes with contrasting views on how agents might change the web. These are early predictions from March 2026, not settled descriptions of how the internet must evolve.