‘We probably have 20 months’ to rebuild AI agents, Meta infrastructure VP tells VB Transform 2026

Organizations need to evolve to meet the needs of agent AI.

Meta VP of Engineering Barak Yagour opened his speech at VB Transform 2026 wearing Ray-Ban Meta AI glasses, a small sign of how far AI has come into physical health. His argument went further: business infrastructure is built for people, not agents, and it’s starting to show.

Yagour, who leads the data infrastructure organization, told the audience that inquiries about Meta’s data systems grew 30x in one half, a turnaround that he said defied expectations the company had spent two decades building.

Change isn’t just limited to Meta. Robotics overtook human traffic on the Internet last year, accounting for 51% of the total, according to Imperva’s 2025 Bad Bot Report. That traffic is growing nearly eight times faster than human traffic, according to HUMAN Security’s 2026 State of AI Traffic report. Yagour cited both figures to describe what he called an inflection point already underway within his organization.

Yagour positioned the change as an open question for infrastructure teams everywhere. "What happened to the infrastructure we spent years building where agents and not people became the main buyers of that," Yagour said. "That is the world we are entering."

Power, identity and speed are broken at the same time

Yagour said three considerations are broken simultaneously within the Meta infrastructure: capacity, ownership and speed.

With volume, analytics no longer work the way engineering teams used to. "One engineer used to mean one unit of load," he said. "Now one engineer produces 10 agents, each of which produces subagents. Your organization of 1,000 people can generate a load of 100,000 users overnight."

His answer is not to block agent traffic but to make the infrastructure agent aware, with dynamic controls that understand the agent’s sequence, cost annotation that tracks back-to-back use of the use case it created, and adaptive motion based on priority.

Identity breaks down, too. Yagour said the agent is not equivalent to the infrastructure layers that access controls are built around. It is not a human user, it does not carry a badge and it is not a service used, yet it makes decisions on its own.

Speed ​​is the third predictor under difficulty. Yagour cited a figure reported by the company that GitHub Copilot writes 46% of the user’s standard code, and noted that faster code generation does not make the rest of the pipeline faster.

"That code still needs to be built, tested, deployed, monitored," he said. "An agent writes code in seconds, but your CI/CD pipeline isn’t faster just because the machine is the writer."

Reliable data states keep agents within monitors

Data is where Yagour said the pressure from agents is most direct.

"Data resides at the center of everything," he said, pointing to the decisions, products, recommendation programs and next-generation models that drive them.

Meta is also rethinking how much autonomy to give agents within its data systems. In February, the company shipped what Yagour called agent data applications. In three months, 63% of dashboards published across Meta were built using the new tools, which is part of the same 30x increase in agent queries Yagour cited earlier.

That growth raises the question of governance. Human analysts have often sat between raw data and business decisions, selecting it and acting as an informal quality check. Yagour said Meta wanted to give agents more freedom on difficult problems, but was specific about the risks.

"Independence without governance is nothing but chaos," he said. That’s why the company has built what it calls trusted data centers, to keep a human check as agents take on more of that work.

"Internally, the agent can freely explore the data, but all output is traced back to its source and processed. So you always know that data is shared and trusted and governed," Yagour said.

Sensitive fields are hidden before the agent can access them, and every access request is checked in real time against what the agent is trying to access, why and whether it’s allowed. Yagour summarized the evaluation method as being more exploratory while less liberating.

Conceptual models rewrite the data layer

Meta models also demand more from the data as they move from alignment to inference.

"Thinking is hungry for data," Yagour said.

Pattern matching works on a few, summarized signals. Thinking requires a full history of behavior, all interactions everywhere over time. Yagour pointed to two shifts already underway within Meta’s infrastructure to keep it going.

Real-time streaming replaces a lot of ETL with scaling pipelines. A pipeline that takes 24 hours to run doesn’t work if the model thinks about the user’s current intent. Yagour said real-time streaming, not batch extract-transform-load processing, becomes the backbone of Meta’s standard programs and recommendations.

Storage is becoming more schema-aware to stop GPU starvation. Meta previously stored user data as opaque blobs without knowing what the data contained, which Yagour said led to large consumption and idle GPU power. The company is now building a storage that understands what it is holding, pulling only the columns and time range for the needs of a particular query. Yagour said Meta is generating 500 million queries per second and a petabyte per second of training data throughput.

That data goes directly into how Meta recommendations behave. Yagour said 42% of Instagram users told the company they wanted to change the algorithm, not fix a single session or setting. Meta feedback is what Yagour calls fully conversational recommendations, where the user tells the system what they’re most interested in and reasons about intent rather than matching keywords. Yagour said the same search term, soccer, will return different results for the average fan looking for highlights than for a club athlete looking for practice, because the system will show who’s asking.

Yagour described the three strands of his discourse, agents, data and recommendations, as reinforcing each other rather than moving independently.

"Agents make data more accessible. Better data makes assumptions. Thinking creates new demands that push agents and infrastructure forward," he said. "This is not a line; is the flywheel."

During the Q&A, an audience member asked if Meta’s push into smarter infrastructure signals the end of traditional file systems in favor of new neural storage methods, and if agents will continue to use SQL as a data interface the way humans do. Yagour said Meta is experimenting on all levels, including asking whether SQL is the right interface for agents at all, and that storage at Meta’s scale is already operating in the multi-digit exabyte range and needs to continue to expand.

Yagour closed his talk with a timeline he believes the industry is working towards. "We have spent 20 years building infrastructure for people. We have about 20 months to rebuild the entire world where people and agents co-create at a high level," Yagour said. "The window is open, but it won’t stay open for long."