Organizations must transform to meet their agent AI needs.
Barak Yagour, Vice President of Meta Engineering, opened his speech VB Transform 2026 wearing a pair of Ray-Ban Meta AI glasses is just a small hint of how much AI is at work in the physical world. His argument went even further: enterprise infrastructure is built for people, not agents, and it’s starting to show.
Yagour, who heads the data infrastructure organization, told the audience that agent queries hitting Meta’s data systems have increased by a factor of 30 in a half, which he says has shattered assumptions the company has built over two decades.
Queue is not limited to Meta. Automated traffic surpassed human traffic on the internet last year, reaching 51% of total traffic Imperva’s 2025 Bad Bot Report. This traffic is also growing about eight times faster than human traffic HUMAN Security’s State of AI Traffic 2026 report. Yagour cited both figures to describe what he called a turning point already underway at his organization.
Yagour framed the change as an open question for infrastructure teams everywhere. "What happens when agents are not the primary consumers of the infrastructure we’ve been building for years?" Yagour said. "This is the world we step into."
Capacity, personality and speed are broken in an instant
Yagour said three assumptions are simultaneously being violated in Meta’s infrastructure: capacity, identity and speed.
Because of its power, math no longer works the way engineering teams are used to. "An engineer envisioned a unit of load," he said. "An engineer now spawns 10 agents, each of which spawns subagents. Your organization of 1,000 people can generate the load of 100,000 users almost overnight."
Its answer is not to block agent traffic, but to make the infrastructure aware of the agent with dynamic controls that understand agent hierarchies, cost attribution that tracks consumption back to the use case that generated it, and priority-based adaptive regulation.
Personality is also disturbed. Yagour said that an agent does not fit into the categories where infrastructure groups build access control around. It’s not a human user, it doesn’t carry a badge, and it’s not a hosted service, but it makes decisions on its own.
Speed is the third assumption under stress. Yagour cited a figure reported by the company that GitHub Copilot’s average user writes 46% of the code, then noted that generating faster code doesn’t necessarily make the rest of the pipeline faster.
"This code is still being built, tested, deployed, monitored," he said. "The agent writes code in seconds, but your CI/CD pipeline doesn’t go any faster just because the machine is the author."
Trusted data environments keep agents inside guards
Yagour said the agents’ pressure was the most direct.
"Data sits at the center of everything," he said, pointing to the decisions, products, recommendation systems and next-generation models he’s driving.
Meta is also rethinking how much autonomy to give agents in its data systems. In February, the company shipped its agent data software, which it calls Yagour. In three months, 63% of dashboards published on Meta were built using the new tools, which is part of the same 30x increase in agent inquiries that Yagour previously noted.
This growth raises the question of management. Human analysts have traditionally sat between raw data and business decisions, curating them and serving as informal quality checks. Yagour said Meta wants to give agents more independence on more difficult problems, but is directly related to risk.
"Autonomy without governance is nothing but chaos," he said. That’s why the company has built what it calls trusted data environments to protect human verification as agents take on more of this work.
"Internally, the agent can explore the information freely, but each output goes back to the source and is carefully checked. So you always know the data being shared back is secure and controlled," Yagour said.
Sensitive areas are masked before the agent reaches them, and each access request is evaluated in real-time against what the agent is trying to access, why, and whether it is allowed. Yagour summarized the approach as exploring broadly and leaving narrowly.
Thinking models rewrite the information layer
As metamodels move from correlation to inference, they require more from the data.
"Thought is hungry for information," Yagour said.
Pattern matching works on sparse, generalized signals. Judgment requires a complete history of behavior, every interaction on every surface over time. Yagour already pointed to two queues in Meta’s infrastructure to keep up.
Replaces batch ETL for real-time streaming pipeline sequencing. A pipeline that takes 24 hours to run when a model thinks about the user’s current intent is not reliable. Real-time streaming, rather than bulk extract-transform-load processing, has become the backbone of Meta’s ranking and recommendation systems, Yagour said.
Memory becomes scheme aware to stop GPU starvation. Meta previously stored user data in opaque blobs without knowing what the data consisted of, which Yagour said led to heavy overhead and wasted GPU capacity. The company now builds memory that understands what it’s storing, fetching only the columns and time ranges that a given query needs. Yagour said Meta has built in a transfer rate of one petabyte per second for 500 million queries per second and training data reads.
This data feeds directly into how Meta’s recommendation systems behave. Yagour said that 42% of Instagram users told the company they wanted to fundamentally change the algorithm, not adjust a single session or setting. Meta’s answer is what Yagour calls fully conversational recommendations, where the user tells the system more of what they want and is more about intent than matching keywords. Yagour said the same search term, soccer, would yield different results for a casual fan looking for highlights than for a club athlete looking for training drills, because the system would justify which one asked.
Yagour described the three themes of his speech, agents, information and recommendations, as reinforcing each other rather than acting independently.
"Agents make data more accessible. Better data justifies. Thinking agents and infrastructure create new requirements that drive forward," he said. "It is not linear; this is the flywheel."
During the Q&A, an audience member asked whether Meta’s push toward smarter infrastructure signals the end of traditional file systems in favor of new neural storage approaches, and whether agents will continue to use SQL as an interface to data just like humans. Yagour said Meta is testing at every level, including whether SQL is the right interface for agents, and that Meta-scale storage already runs in the multi-digit exabyte range and continues to expand.
Yagour closed his talk with a timeline of what he believes the industry is working against. "We have spent 20 years building infrastructure for people. We have maybe 20 months to rebuild everything for a world co-created by humans and agents at scale." Yagour said. "The window is open, but it won’t stay open for long."
