Speaking at VB Transform 2026, Meta's VP of Engineering for Data Infrastructure Barak Yagour issued a stark warning: enterprise infrastructure, built for human users, must be fundamentally redesigned within 20 months to accommodate AI agents as primary consumers. Wearing Ray-Ban Meta AI glasses, Yagour symbolised how far AI has already integrated into physical life, but quickly stressed that digital foundations are not ready for this transition.
Agent queries surged 3000% in six months
Yagour revealed that agentic queries hitting Meta's data systems grew 30x in a single half-year, shattering assumptions the company had built over two decades. This is not an isolated trend. Automated traffic surpassed human internet traffic last year, reaching 51% of total traffic according to Imperva's 2025 Bad Bot Report, and grows eight times faster than human traffic, per HUMAN Security's 2026 report. Yagour called it an inflection point already underway.
Sponsored Protocol
Capacity, identity and velocity break simultaneously
According to Yagour, three traditional pillars are collapsing: capacity, identity and velocity. On capacity, one engineer now spawns 10 agents each spawning subagents, generating the load of 100,000 users overnight. The answer is not to block traffic but to build agent-aware infrastructure with dynamic controls, cost attribution and priority-based throttling. Identity is problematic because an agent is neither a human user nor a deployed service, yet it makes autonomous decisions. Velocity faces strain because tools like GitHub Copilot write 46% of average code, but CI/CD pipelines do not speed up accordingly. Faster code generation does not accelerate build, test, deploy and monitor cycles.
Trusted data environments govern autonomy
Data sits at the heart of the pressure, Yagour said. Meta launched agentic data apps in February, and within three months, 63% of published dashboards used the new tooling. Autonomy without governance is chaos, so Meta built trusted data environments where agents explore freely but every output is traced and scrutinized, sensitive fields are masked, and access requests are evaluated in real time. Yagour summarized: explore broadly, release narrowly.
Sponsored Protocol
Reasoning models rewrite the data layer
The shift from pattern matching to reasoning demands full behavioral history, not sparse signals. To support this, Meta is replacing batch ETL with real-time streaming for ranking pipelines and building schema-aware storage to prevent overfetching and GPU starvation. The target is 500 million queries per second and a petabyte per second throughput for training data reads. These systems feed conversational recommendations: 42% of Instagram users want fundamental algorithmic changes, and Meta is responding with fully conversational recommendations that reason about intent, not just keywords.
Sponsored Protocol
A 20-month window for the industry
Yagour closed with a stark timeline: "We spent 20 years building infrastructure 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 will not stay open long. As companies like Meta invest in agent-aware infrastructure, the industry faces radical choices. For more context, read about how an ex-Apple engineer raised $52 million for Hemispheric, an AI brain diagnosis startup. Also, the UK government proposes a social media curfew for teens, highlighting regulatory responses to AI-driven changes. For a broader perspective, see Wikipedia's entry on intelligent agents.
