Most enterprise AI discussions start with models, APIs, and latency. Arun Joseph's presentation on agentic compute flips that script, grounding the conversation in something far less glamorous but far more consequential: organizational fault lines. Drawing on Deutsche Telekom's LMOS platform, Joseph argues that the real barrier to scaling AI isn't model quality but the messy reality of how enterprises actually operate. That means fragmented tools, competing team priorities, and a stubborn gap between what leadership wants and what engineers can safely deploy. It's a refreshingly honest framing, especially when so much of the industry is still chasing demos that fall apart outside a slide deck.
The throughline here is abstraction. Joseph isn't advocating for more connectors or a bigger pile of integrations. He's making the case for replacing tool sprawl with a small set of core platform abstractions, and then letting ephemeral agents do the heavy lifting. This is where the presentation connects to broader conversations about Graph Engineering for AI Agents: From Prompts and Loops to Workflows. Teams get stuck debating loops versus graphs when the real question is how to impose structure without losing flexibility. Joseph's Agent Definition Language (ADL) is an attempt to answer that question directly. It's a way of encoding agent behavior that's explicit, reviewable, and, critically, aligned with how enterprises already think about governance and change management.
What makes this more than a technical talk is the emphasis on moving beyond chatbots. Joseph isn't interested in building a better Q&A widget. He's describing operational intelligence systems that can sense, reason, and act across an enterprise. That's a meaningful ambition, but it also raises a question that Decision Models in Agentic Architectures: From Production to Agent Skills tackles head-on: how do you trust non-deterministic systems in high-stakes environments? Joseph's answer, at least implicitly, is that you design for accountability from the start. You don't bolt on safety later. You build it into the language, the platform, and the organizational workflow. That's a harder path, but it's the only one that leads anywhere durable.
If there's a takeaway worth quoting, it's this: the future of enterprise AI belongs to teams that treat agentic systems as infrastructure, not experiments. That means investing in the unglamorous work of defining interfaces, setting boundaries, and getting different parts of the organization to actually talk to each other. It's not about being first. It's about being deliberate. And as Explore the Future of AI Deployment: Key Topics at QCon AI New York suggests, those conversations are only going to get more urgent. The specific detail to watch is whether ADL or something like it gains traction beyond a single telecom giant. Because if it does, we're not just looking at a better tool. We're looking at a new standard for how enterprises build, deploy, and trust AI systems in production.