Agentic Infrastructure Stack Summit

The real test of enterprise AI isn’t just what the technology can do, but whether it can deliver better customer experiences reliably and at scale. For T-Mobile, that means moving beyond standalone AI capabilities to building the infrastructure that connects AI models, enterprise systems and customer interactions.

In this fireside chat, T-Mobile CIO Jeff Simon will share how the company is bringing AI into real customer experiences across digital and care, from conversational AI that helps customers resolve issues to Journey Bot in T-Life, which guides customers through tasks like upgrading their devices. With more than 30 million monthly active T-Life users and AI increasingly embedded across customer journeys, T-Mobile is demonstrating what it takes to deliver AI-powered experiences at enterprise scale.

The conversation will go behind the scenes of the technology making this possible, including how T-Mobile orchestrates multiple AI agents, connects models to customer and account information, manages context and tool calls, and balances performance, reliability and efficiency. Jeff will also share lessons learned along the way, from determining when to use specialized agents to ensuring AI has the right information to deliver accurate, useful responses.

Ultimately, the discussion will explore what it takes to turn AI capabilities into meaningful customer outcomes, and why the real differentiation comes from how enterprises bring together AI, technology and people to deliver better experiences.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

This block of the summit runs under the line “Fewer agents, better coordinated.” WEKA chief AI officer Val Bercovici argues the opposite. Jobs keep splitting into more subtasks, agents increasingly hand work to other agents, including agents at other companies, and security teams now want agents running around the clock instead of on demand. In Bercovici’s view, that growth makes memory the limit on orchestration. Every agent rereads the same documents, history and instructions before it can work, and that rereading can outweigh the work itself: In coding sessions SemiAnalysis measured in August, the median request carried 88,000 tokens of input to produce 413 tokens of output. In this conversation, Bercovici explains why demand for agents keeps rising even as enterprises cut back on token spending, how to coordinate many agents around shared memory so each one picks up context instead of starting over, and what infrastructure teams should have in place before their agent count climbs.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

PayPal runs AI agents that write code, automate internal workflows, diagnose production incidents, and help shoppers complete their purchases. To support agents at enterprise scale, PayPal is building a unified platform with certified AI asset registries, a gateway for governance and controls, enterprise memory, and a scalable runtime that supports multiple agent frameworks. In this conversation, Deepika Khera, Senior Director of AI at PayPal, explains how PayPal decides how many agents a workflow needs, how enterprise memory helps agents improve with every task, and how a governed platform with compliance built in improves performance, lowers the cost of AI tasks and measures the business value AI delivers.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

Your AI agents are only as intelligent as the context they can access—and only as trustworthy as the data infrastructure behind them.

As enterprises move from standalone AI assistants to interconnected, multi-agent systems, a new infrastructure challenge is emerging: how do you manage persistent memory, shared context, and real-time retrieval across thousands of agents without compromising security, compliance, or performance? And with agentic technologies evolving at breakneck speed, how do you architect systems that won’t be obsolete in six months?

In this provocative fireside chat, Ben Kus, CTO of Box, joins Matt Marshall, Founder and Editor-in-Chief of VentureBeat, to challenge conventional thinking about enterprise AI architecture. Drawing on Box’s experience building enterprise-scale AI capabilities, they’ll explore how organizations can design flexible, model-agnostic memory and context infrastructure that supports multi-agent collaboration while navigating data governance, access controls, GDPR compliance, and the growing complexity of agent-to-agent interactions.

From persistent memory and intelligent retrieval to cross-platform interoperability and rapidly evolving model ecosystems, this conversation will examine the architectural trade-offs that matter most.

The central question: How do you build enterprise AI infrastructure that remembers everything it needs to, protects everything it must, and adapts to whatever comes next?

Moderated by Matt Marshall, CEO & Founder, VentureBeat

As AI agents move from pilots into production, organizations are discovering that traditional control models can slow adoption while failing to reveal distributed, changing risk. This roundtable will explore how senior technology and AI leaders can build practical guardrails that preserve speed, accountability, and visibility at scale. What if the goal is not to control every agent, but to decide where governance matters most? We will unpack decision-centered oversight, people-process-platform alignment, federated operating models, continuous evidence, and maturity measures that turn policy into action. Attendees will leave with sharper questions, peer-tested perspectives, and concrete moves for governing agents in production.

Moderated by Rob Strechay, Lead Analyst, VentureBeat

The memory and context sessions at this summit start with a premise: less context produces better answers. Anjul Bhambhri, senior vice president of engineering at Adobe, argues that it’s not so much about “less” vs. “more.” Instead, it’s about optimizing and right-sizing context. She explores this topic through the lens of Adobe CX Coworker, an agentic platform purpose-built for marketing and customer experience teams. Bhambhri explains how Coworker builds context that is deliberately narrow and opinionated. AI agents in marketing face an ‘ocean’ of information: documents like campaign briefs and brand guidelines, data dictionaries that define the business terms, and live systems of record such as the CDP, CRM and ERP. Instead of trying to boil the ocean, Coworker uses techniques such as distilling information in advance, leveraging knowledge graphs, indexing, merging and ranking results to construct the right context at the right time. Right-sizing context not only reduces token costs but also improves outcomes.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

As AI costs climb, the enterprises winning with agents are the ones who get more done per task, not the ones running the biggest models. That shift is moving attention from the model to the memory and context layer, and turning agentic AI back into a search problem. In this session, Databricks’ Bryan Clark will explain why a dedicated vector database is no longer the default, and how a serverless Postgres database built for agents can combine agent state, hybrid retrieval and fresh operational data in one place. He’ll share the emerging enterprise playbook: Start with frontier models, make the infrastructure smarter, then switch to cheaper open-weight models without sacrificing accuracy.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

A 2025 UC Berkeley study of seven open-source multi-agent systems found they failed on 41% to 87% of their tasks, and that about a third of the failures it catalogued came from agents misaligning with one another, not from mistakes inside any single agent. The pattern is familiar to anyone building real systems: the agents connect, the messages move, the tools fire, and the outcome still breaks down.

Connectivity was never the hard part. Enterprise agents can already discover one another, exchange messages, and invoke the same tools. What’s missing is semantic alignment, the layer where agents verify what they’re actually solving for before they act, not just pass messages back and forth. Context is not coordination.

In production, that gap turns into real risk when an agent acts on a task it never correctly understood. This conversation is about closing that orchestration gap, grounded in real deployments, and what infrastructure teams must build next before “agentic” becomes an operating model instead of a demo.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

Based on Artificial Analysis’s independent benchmarking, the talk explores how model intelligence, token consumption, and inference metrics determine cost and latency for real AI tasks, along with tradeoffs between frontier and open weights models and model-routing strategies.

Moderated by Rob Strechay, Lead Analyst, VentureBeat

Uber operates one of the world’s largest mobility platforms, powering more than a billion trips and deliveries every month, with thousands of engineers working across 12 global sites. As AI agents take on an increasing share of code generation, code review, CI repair, on-call triage, and code maintenance, the challenge is no longer simply automating software development—it is making agentic engineering economically viable at scale.

This session goes under the hood of Uber’s AI implementations to explore the architecture and engineering decisions that determine the economics of agentic workloads. From model routing, prompt caching, and token efficiency to inference benchmarking, context management, and orchestration, the discussion will examine how to optimize the trade-offs between model capability, latency, reliability, and cost. It will also explore how evaluation frameworks and human escalation affect the economics of autonomous workflows, and how engineering teams can measure success beyond raw token costs to understand the cost per successful task. As agent adoption accelerates, what does it take to scale AI-driven software development without letting inference costs outpace productivity gains?

Moderated by Rob Strechay, Lead Analyst, VentureBeat

Token prices keep falling, and enterprises keep spending more on AI agents anyway. Model usage comes out of one budget, GPUs out of another and engineering time out of a third, so few enterprises can say what a single agent costs to finish its work. Many finance chiefs are now handing business teams a fixed budget and telling them to manage the tokens themselves. This panel brings together Sudhanshu Bist, vice president of platform infrastructure engineering at Upwork, and Val Bercovici, chief AI officer of WEKA, to move the conversation from token economics to task economics: how to measure the full cost of an agent against the work it completes, which numbers to track before scaling, and why the option that looks cheapest often costs the most.

Moderated by Matt Marshall, CEO & Founder, VentureBeat

As AI evolves from isolated model inference to autonomous agents that reason, plan, and execute complex workflows, the infrastructure powering AI must evolve with it. What does an AI factory built for the agentic era actually look like?

In this fireside chat, Reynold D’Sa, Corporate Vice President of Microsoft’s Silicon, Cloud Hardware, and Infrastructure Engineering organization, takes us inside the emerging architectures reshaping AI infrastructure at hyperscale. From custom silicon and accelerated compute to high-performance networking, memory, and intelligent resource orchestration, the discussion explores how Microsoft is co-designing the AI factory across the infrastructure stack for increasingly dynamic, compute-intensive agentic workloads. We’ll also examine how lessons from operating AI at global scale, and applying AI to infrastructure engineering itself, are shaping the next generation of AI factories, where performance, efficiency, and scalability advance together.

Moderated by Rob Strechay, Lead Analyst, VentureBeat

Moving AI agents into production takes more than models and compute. It requires infrastructure that governs how agents access models, tools, and data. Join Tetrate co-founder and CTO Varun Talwar for a discussion on where AI gateways fit in the AI factory, how to assess the maturity of open-source building blocks like Agent Router, and which architectural choices preserve flexibility as the ecosystem evolves.

Moderated by Rob Strechay, Lead Analyst, VentureBeat

The next frontier of agentic AI isn’t building more agents. It’s building the infrastructure that allows them to work together. As enterprises move from experimentation to production, the real challenge is connecting agents, data and existing systems without sacrificing control, security or flexibility. In this conversation, technology leaders from Omnicom’s Omni and Acxiom join VentureBeat to explore what it takes to build an open, interoperable foundation—one where intelligence moves across systems, agents operate within connected workflows, and governance is embedded at every step. They’ll examine the architectural decisions that matter most and why the enterprises that get the infrastructure right today will have a distinct advantage tomorrow.

Moderated by Rob Strechay, Lead Analyst, VentureBeat

Anna Nicanorova is a technology leader at Omnicom, the world’s leading marketing and communications company. Since 2018, she has played a key role in the development and evolution of Omni, Omnicom’s connected intelligence platform—driving its technology architecture and growth strategy.

As Chief Technology Officer at Omnicom-owned Annalect, Anna expanded Omni to hundreds of global users through a modernized data mesh infrastructure. She also led the development of privacy-compliant technologies such as clean rooms and spearheaded system-wide improvements in modular, high-performance enterprise software.

Anna's work in emerging technologies—including machine learning, augmented reality, and computer vision—has been applied across marketing and featured in publications such as CNBC, Adweek, and Ad Age. She has been named to Ad Age’s “40 Under 40,” received Starburst’s Executive AI & Data Visionary Award and recently served as a Cannes Lions Innovation juror.

Anna holds a BA from Hogeschool Utrecht University of Applied Sciences and an MBA from Wharton. Outside of work, she volunteers as a coding instructor in under-resourced schools, climbs mountains around the world, and is training for her pilot’s license.