The Infrastructure Bottleneck

Data transfer costs are often invisible, yet they are rapidly becoming the most significant variable eroding enterprise infrastructure budgets. Barak Yagour, Vice President of Engineering at Meta, recently warned at VB Transform 2026 that existing infrastructure designs are fundamentally incapable of handling the explosive growth of agent-based traffic. According to Yagour, organizations have a window of approximately 20 months to re-architect their systems before current data processing methods become a catastrophic bottleneck. In an environment where autonomous agents, rather than humans, drive the majority of traffic, the legacy approaches to data retrieval are failing to keep pace.

Transitioning to Schema-Aware Storage

Meta is currently pivoting its data storage strategy toward schema-aware structures to maximize GPU efficiency. Historically, data was stored as opaque blobs, a method that prevented systems from understanding the underlying data characteristics. This lack of visibility led to excessive data overfetching and significant GPU idle time, as systems repeatedly pulled unnecessary information. To combat this, Meta is implementing an architecture that precisely extracts only the specific columns and time ranges required for a given query. This shift is a mandatory evolution to minimize the GPU waste that occurs when autonomous agents perform data lookups and code generation tasks.

Balancing Autonomy with Governance

Redesigning data pipelines is not merely a performance optimization; it is a fundamental rewrite of the rules governing how agents interact with sensitive information. Meta has established trusted data environments to maintain strict governance while enabling agent autonomy. Within these environments, agents can explore data freely, but every output is subject to rigorous source tracking and validation. Sensitive fields are masked before an agent can access them, and all data requests are evaluated in real-time to ensure security protocols remain intact.

Furthermore, the traditional batch ETL process, which often requires a 24-hour cycle, is increasingly incompatible with inference models that must understand a user's immediate intent. Meta is transitioning its infrastructure from batch processing to real-time streaming. This shift is now the cornerstone of their ranking and recommendation systems, ensuring that agents can make accurate decisions based on live data rather than stale, pre-processed snapshots.

The Shift in Query Patterns

Internal metrics shared by Yagour highlight the urgency of this transition: agentic queries—requests where AI agents autonomously query and analyze data—have surged 30-fold in just six months. This velocity is dismantling the core assumptions of infrastructure design that have been optimized for human user patterns over the past two decades. This phenomenon mirrors the broader internet trend where automated traffic has overtaken human activity, forcing a complete rethink of operational foundations.

As 42% of Instagram users demand fundamental changes to algorithmic experiences, Meta is deploying fully conversational recommendation systems. These systems move beyond keyword matching, allowing the engine to infer user intent directly. As users and agents begin to interact with databases through natural language, the reliance on static data processing is becoming a liability. The transition to real-time streaming and schema-aware storage is no longer optional; it is the only way to eliminate GPU idle time and satisfy the demands of agentic workflows.

With agentic queries set to become the primary driver of infrastructure costs over the next 20 months, engineering teams must immediately verify whether their current data governance frameworks can maintain control within a real-time streaming environment.