Organizations want to remodel to satisfy the wants of agentic AI.
Meta VP of Engineering Barak Yagour opened his speak at VB Remodel 2026 sporting a pair of Ray-Ban Meta AI glasses, a small signal of how far AI has already labored its approach into bodily life. His argument went additional: enterprise infrastructure was constructed for people, not for brokers, and it's beginning to present.
Yagour, who leads its knowledge infrastructure group, advised the viewers that agentic queries hitting Meta's knowledge techniques grew 30x in a single half, an inversion that he mentioned is breaking assumptions the corporate spent 20 years constructing round.
The shift is just not confined to Meta. Automated site visitors overtook human site visitors on the web final yr, reaching 51% of the whole, in keeping with Imperva's 2025 Unhealthy Bot Report. That site visitors can be rising roughly eight occasions quicker than human site visitors, in keeping with HUMAN Safety's 2026 State of AI Site visitors report. Yagour cited each figures to explain what he referred to as an inflection level already underway inside his personal group.
Yagour framed the shift as an open query for infrastructure groups all over the place. "What occurs to the infrastructure we've spent years constructing when brokers and never people grow to be the principle customers of that," Yagour mentioned. "That's the world we're moving into."
Capability, id and velocity are breaking without delay
Yagour mentioned three assumptions are breaking concurrently inside Meta's infrastructure: capability, id and velocity.
On capability, the maths not works the best way engineering groups are used to. "One engineer used to imply one unit of load," he mentioned. "Now one engineer spawns 10 brokers, every spawning subagents. Your 1,000-person org can generate the load of 100,000 customers virtually in a single day."
His reply is to not block agent site visitors however to make infrastructure agent-aware, with dynamic controls that perceive agent hierarchies, value attribution that traces consumption again to the use case that spawned it, and throttling that adapts primarily based on precedence.
Id is breaking, too. Yagour mentioned an agent doesn’t match the classes infrastructure groups constructed entry controls round. It isn’t a human consumer, it doesn’t carry a badge and it’s not a deployed service, but it makes choices by itself.
Velocity is the third assumption beneath pressure. Yagour cited a company-reported determine that GitHub Copilot writes 46% of the common consumer's code, then famous that quicker code technology doesn’t make the remainder of the pipeline quicker.
"That code nonetheless must be constructed, examined, deployed, monitored," he mentioned. "The agent writes the code in seconds, however your CI/CD pipeline doesn't get quicker simply because the machine is the writer."
Trusted knowledge environments preserve brokers inside guardrails
Information is the place Yagour mentioned the strain from brokers is most direct.
"Information sits on the middle of every part," he mentioned, pointing to the selections, merchandise, recommender techniques and subsequent technology fashions it drives.
Meta can be rethinking how a lot autonomy to grant brokers inside its personal knowledge techniques. In February, the corporate shipped what Yagour referred to as agentic knowledge apps. Inside three months, 63% of dashboards printed throughout Meta had been constructed utilizing the brand new tooling, a part of the identical 30x rise in agentic queries Yagour cited earlier.
That progress raises a governance query. Human analysts have historically sat between uncooked knowledge and enterprise choices, curating it and serving as a casual examine on high quality. Yagour mentioned Meta needs to grant brokers extra independence on tougher issues, however was direct in regards to the threat.
"Autonomy with out governance is nothing however chaos," he mentioned. That's why the corporate constructed what it calls trusted knowledge environments, to protect the human examine as brokers tackle extra of that work.
"Inside, the agent can discover knowledge freely, however each output is traced again to its supply and scrutinized. So that you all the time know that the info shared again is trusted and ruled," Yagour mentioned.
Delicate fields are masked earlier than an agent can attain them, and each entry request is evaluated in actual time in opposition to what the agent is making an attempt to achieve, why and whether or not it’s allowed. Yagour summarized the strategy as exploring broadly whereas releasing narrowly.
Reasoning fashions are rewriting the info layer
Meta's fashions are additionally demanding extra from knowledge as they shift from correlation to reasoning.
"Reasoning is knowledge hungry," Yagour mentioned.
Sample matching works on sparse, summarized alerts. Reasoning calls for the total behavioral historical past, each interplay throughout each floor over time. Yagour pointed to 2 shifts already underway inside Meta's infrastructure to maintain up.
Actual-time streaming is changing batch ETL for rating pipelines. A pipeline that takes 24 hours to run is just not viable when a mannequin is reasoning a couple of consumer's present intent. Yagour mentioned real-time streaming, not batch extract-transform-load processing, is turning into the spine of Meta's rating and suggestion techniques.
Storage is turning into schema-aware to cease GPU hunger. Meta beforehand saved consumer knowledge as opaque blobs with no consciousness of what the info contained, which Yagour mentioned led to heavy overfetching and idle GPU capability. The corporate is now constructing storage that understands what it holds, pulling solely the columns and time ranges a given question wants. Yagour mentioned Meta is constructing towards 500 million queries per second and a petabyte per second of throughput for coaching knowledge reads.
That knowledge feeds straight into how Meta's suggestion techniques behave. Yagour mentioned 42% of Instagram customers have advised the corporate they wish to essentially change the algorithm, not regulate a single session or setting. Meta's response is what Yagour referred to as totally conversational suggestions, the place a consumer tells the system what they need extra of and it causes about intent quite than matching on key phrases. Yagour mentioned the identical search time period, soccer, would return totally different outcomes for an off-the-cuff fan in search of highlights than for a membership athlete in search of coaching drills, as a result of the system would purpose about which one is asking.
Yagour described the three threads of his speak, brokers, knowledge and proposals, as reinforcing one another quite than shifting independently.
"Brokers make knowledge extra accessible. Higher knowledge makes reasoning. Reasoning creates new calls for that push brokers and infrastructure ahead," he mentioned. "This isn't linear; it's a flywheel."
In the course of the Q&A, an viewers member requested whether or not Meta's push towards extra clever infrastructure alerts the top of conventional file techniques in favor of newer neural storage approaches, and whether or not brokers will preserve utilizing SQL as their interface to knowledge the best way people do. Yagour mentioned Meta is experimenting at each stage, together with questioning whether or not SQL is the correct interface for brokers in any respect, and that storage at Meta's scale already operates within the multi-digit exabyte vary and must preserve increasing.
Yagour closed his speak with the timeline he believes the business is working in opposition to. "We spent 20 years constructing infrastructure for people. We now have possibly 20 months to rebuild the entire thing for a world the place people and brokers co-create at scale," Yagour mentioned. "The window is open, nevertheless it gained't keep open for lengthy."
