Barak Yagour, Vice President of Engineering responsible for data infrastructure at Meta, told VB Transform 2026 that enterprise systems were designed for human users but are now being stressed by agentic (agent‑driven) traffic. He opened his talk wearing Ray‑Ban Meta AI glasses as a visual cue for how AI is already present in physical life.
What happened and why it matters
Yagour said agentic queries hitting Meta’s data systems rose 30x within a single half‑year. External reports mirror the trend: Imperva’s 2025 Bad Bot Report found automated traffic exceeded human traffic on the internet last year (51%), and HUMAN Security’s 2026 State of AI Traffic report shows automated traffic growing roughly eight times faster than human traffic.
According to Yagour, this shift is challenging three core infrastructure assumptions at once: capacity, identity and velocity.
Capacity: one engineer no longer equals one unit of load
Traditional capacity assumptions are breaking because one engineer can today spawn multiple agents and subagents. "One engineer used to mean one unit of load," Yagour said. "Now one engineer spawns 10 agents, each spawning subagents. Your 1,000‑person org can generate the load of 100,000 users practically overnight."
Meta’s answer is not to block agent traffic but to make infrastructure agent‑aware: dynamic controls that understand agent hierarchies, cost attribution that traces consumption back to the originating use case, and priority‑based throttling.
Identity: agents don’t fit existing access models
An agent is not a human user, does not carry a badge, and is not necessarily a deployed service, yet it makes autonomous decisions. That mismatch complicates access control, authentication and audit trails.
Velocity: faster generation doesn’t speed the whole pipeline
Yagour cited a company figure that GitHub Copilot writes 46% of the average user’s code, but faster code generation doesn’t make builds, tests, deployment or monitoring faster. "The agent writes the code in seconds, but your CI/CD pipeline doesn't get faster just because the machine is the author," he said.
Data and governance: trusted data environments
Yagour said agents exert the most pressure at the data layer. Meta shipped so‑called agentic data apps in February, and within three months 63% of dashboards published across Meta were created using that tooling—part of the same 30x increase in agentic queries.
That rapid uptake raises governance questions. Historically, human analysts acted as an informal quality check between raw data and business decisions. Meta wants to give agents more autonomy on harder problems but recognizes the risk: "Autonomy without governance is nothing but chaos." To address this, Meta built trusted data environments where agents can explore data while every output is traced and scrutinized. Sensitive fields are masked before agents can reach them, and each access request is evaluated in real time against what the agent is trying to access, why, and whether it's permitted. Yagour summarized the approach as exploring broadly while releasing narrowly.
Reasoning models are changing data requirements
As Meta’s models move from correlation to reasoning, they demand more and different data. "Reasoning is data hungry," Yagour said, because it needs full behavioral histories and every interaction across surfaces over time rather than sparse, summarized signals.
Meta is adapting in two concrete ways:
- Real‑time streaming is replacing batch ETL for ranking pipelines. A pipeline that runs every 24 hours is not viable when a model must reason about a user’s current intent.
- Storage is becoming schema‑aware to avoid GPU starvation. Previously, Meta stored user data as opaque blobs, which led to overfetching and idle GPU capacity. Now storage understands what it contains and pulls only the columns and time ranges required by a query. Yagour said Meta is building toward 500 million queries per second and petabyte‑per‑second throughput for training data reads.
These data changes feed directly into recommendation systems. Yagour said 42% of Instagram users told Meta they would fundamentally change the algorithm rather than tweak a single session or setting. Meta’s response is fully conversational recommendations that reason about intent—for example, the same search term "soccer" should return different results for a casual fan seeking highlights versus a club athlete seeking training drills.
Interconnected trends and experimentation
Yagour framed agents, data and recommendations as reinforcing each other rather than moving independently: "Agents make data more accessible. Better data makes reasoning. Reasoning creates new demands that push agents and infrastructure forward. This isn't linear; it's a flywheel."
In response to an audience question about whether traditional file systems will give way to neural storage and whether agents will keep using SQL as their data interface, Yagour said Meta is experimenting at every level, including whether SQL is the right interface for agents. He noted Meta’s storage already operates in the multi‑digit exabyte range and must continue expanding.
Timeline and takeaway
Yagour closed with a timeline: "We spent 20 years building infrastructure for humans. We have maybe 20 months to rebuild the whole thing for a world where humans and agents co‑create at scale. The window is open, but it won't stay open for long."
For other organizations, the lesson is urgent: agent traffic and capabilities are growing faster than traditional infrastructure changes, so teams should invest in agent‑aware controls, real‑time pipelines, schema‑aware storage and robust governance to keep agents inside guardrails while enabling their utility.
