Meta’s VP of Engineering, Barak Yagour, warned at VB Transform 2026 that enterprise infrastructure is failing under the weight of agentic AI. With agentic queries surging 30x in six months, current systems built for human-centric interaction are obsolete. Organizations have approximately 20 months to rearchitect for an autonomous, agent-led digital environment.
The Structural Collapse of Capacity and Identity
The traditional “one-to-one” model of computing is dead. For two decades, infrastructure teams scaled systems based on the assumption that one human user equals one unit of load. That math no longer holds. As Yagour noted, a single engineer can now spawn ten agents, which in turn spawn sub-agents, creating a recursive load that can overwhelm a 1,000-person organization with the traffic of 100,000 users in hours.
This is not just a capacity issue; it is an identity crisis. Access control lists (ACLs) and identity management systems were designed for human users with badges or service accounts with static permissions. Autonomous agents, which make real-time decisions, do not fit these rigid buckets. Without “agent-aware” infrastructure that understands hierarchies and dynamic cost attribution, enterprises face a future of runaway cloud bills and uncontrollable compute sprawl.
Data Governance in the Age of Autonomy
The shift toward agentic AI forces a reckoning with how data is accessed and curated. Historically, human analysts acted as the “human-in-the-loop” filter, ensuring data quality before it reached the boardroom. With the rise of agentic data apps—which now account for 63% of dashboards at Meta—that human buffer is eroding.
To prevent “autonomy without governance” from descending into chaos, Meta is implementing “trusted data environments.” These systems treat data access not as a static permission but as a real-time negotiation. Every query is scrutinized: what is the agent trying to reach, why is it doing so, and does it align with the business intent? This approach mandates that sensitive fields be masked at the storage layer, ensuring that even if an agent is authorized to “explore broadly,” its output is “released narrowly.”
Why Reasoning Models Break Traditional ETL Pipelines
The industry is moving from simple pattern matching to intensive reasoning, and the infrastructure cost is massive. Reasoning models are “data-hungry” in a way that previous LLMs were not. They require full behavioral histories, not just summarized signals. Consequently, the standard batch-based Extract-Transform-Load (ETL) pipelines are becoming a bottleneck.
Meta is actively replacing these 24-hour batch cycles with real-time streaming to support recommendation engines. Furthermore, the company is abandoning “opaque blob” storage. By moving to schema-aware storage, systems can now pull only the specific columns and time ranges required for a query, drastically reducing GPU starvation. The goal: supporting 500 million queries per second with a throughput of a petabyte per second.
The 20-Month Deadline: Industry Perspectives
The 20-month window Yagour identified is a aggressive timeline for a wholesale infrastructure overhaul. The pressure is compounded by the fact that automated traffic has already crossed the 51% threshold, effectively making the internet a machine-first ecosystem according to Imperva’s 2025 Bad Bot Report. Engineering teams are currently caught between maintaining legacy stability and building for this agent-driven future.
Some industry observers highlight that this shift requires a move beyond SQL as the primary interface for data. Similarly, the GitHub Copilot ecosystem has already demonstrated that while agents can accelerate code generation by nearly 50%, the underlying CI/CD pipelines remain a legacy friction point that requires modernization to match the speed of the machine author.
The 30-Second Verdict
- Capacity: Move from static load balancing to dynamic, agent-aware throttling.
- Identity: Abandon human-centric access controls for context-aware, recursive permissioning.
- Data: Shift from batch ETL to real-time, schema-aware streaming to feed reasoning models.
- Governance: Implement “trusted data environments” where every agent output is traceable to a source.
The transition is not linear; it is a flywheel. Agents make data accessible, better data improves reasoning, and reasoning demands more from the infrastructure. The window to rebuild is open, but for those relying on human-era assumptions, it is closing rapidly.
