Meta’s 20‑Month Sprint to Rebuild Infrastructure for AI agents

AI agents reshape enterprise infrastructure

At VB Transform 2026, Meta’s VP of Engineering Barak Yagour warned that the tech giant has roughly ↑20 months to overhaul its data backbone for AI agents. He opened wearing Meta AI glasses, a visual cue that the technology is already woven into daily life. Yagour cited a ↑30x surge in agentic queries to Meta’s data platforms in just six months, a reversal that shatters two decades of human‑centric design assumptions. Automated traffic now exceeds human traffic worldwide, hitting 51% of all internet requests last year, according to the Imperva Bad Bot Report and echoed in a recent Reuters analysis. The growth rate of bot traffic is eight times faster than that of human users, a trend highlighted by HUMAN Security’s 2026 State of AI Traffic study. Yagour framed the shift as a fundamental question for every infrastructure team: “What happens when agents, not humans, become the primary consumers of the systems we built?” He identified three pillars under pressure: capacity, identity and velocity. Capacity calculations now collapse; a single engineer can spin up ten agents, each spawning sub‑agents, turning a 1,000‑person team into the load of 100,000 users overnight. Instead of blocking agents, Meta is making its stack agent‑aware, adding dynamic controls that trace consumption back to originating use cases and prioritize traffic in real time. Identity models crumble because agents lack badges, human profiles, or static service tags, yet they make autonomous decisions. Velocity is strained as tools like GitHub Copilot now generate 46% of a developer’s code in seconds, but the downstream CI/CD pipeline remains unchanged, creating bottlenecks. Data sits at the epicenter of this upheaval. Meta’s new “trusted data environments” let agents explore data freely while every output is logged, masked where necessary, and vetted against real‑time policy checks. Sensitive fields are hidden, and each request is evaluated for purpose and permission before access is granted. Yagour emphasized that autonomy without governance breeds chaos, so the company embeds traceability into every data transaction. Reasoning‑heavy models now demand full behavioral histories rather than sparse signals, pushing Meta toward real‑time streaming pipelines and schema‑aware storage that fetches only the needed columns, avoiding GPU starvation. The goal: 500 million queries per second and a petabyte‑per‑second throughput for training data. On the consumer side, Meta is testing “fully conversational recommendations” on Instagram, letting users articulate intent—e.g., a casual fan versus an athlete searching “soccer”—and receiving results tailored by reasoning rather than keyword matching. Yagour concluded with a stark timeline: “We spent two decades building for humans. We have maybe 20 months to rebuild for a world where humans and agents co‑create at scale. The window is open, but it won’t stay open for long.”