Within barely twenty months, Meta warns, enterprises must overhaul decades-old infrastructure built for human-centric workloads if they hope to capture value from the coming wave of AI agents. The company’s infrastructure leadership frames this window as an industry-wide inflection point, not a single-platform concern, arguing that systems accumulated over roughly two decades for human users are misaligned with emerging human–agent co-creation at scale. Delay, they contend, risks systemic bottlenecks once autonomous agents become dominant generators of digital traffic.
Meta’s warning is grounded in rapidly escalating agentic demand inside its own data infrastructure. Over a single half, the company reports a thirtyfold increase in agent-originated queries, straining capacity models built around relatively infrequent human actions. Legacy planning assumed that one engineer or one user produced one unit of load; under agentic patterns, each human may launch networks of agents and subagents, multiplying traffic until a thousand-person organization effectively behaves like tens of thousands of users.
Agentic demand explodes: 30x query growth turns thousand-person teams into pseudo populations of tens of thousands
Inside such environments, three long-standing assumptions—capacity, identity, and velocity—are described as breaking simultaneously. Capacity models tuned to steady human usage no longer match continuous, high-frequency agent traffic that operates at machine speed. Identity frameworks built around static user accounts and badges struggle to classify agents that are neither people nor conventional services, yet act autonomously and hold credentials. Velocity expectations based on daily or batch data movements falter when agents demand real-time interactions and low-latency access to operational data. A lack of structured governance in managing these agents further complicates the situation.
To prevent infrastructure from becoming the limiting factor, Meta promotes a shift away from twenty-four-hour batch ETL pipelines toward streaming architectures built for continuous agentic workloads. Future-facing targets include being able to handle hundreds of millions of queries per second and petabyte-per-second training data reads, reflecting a step change in how data is moved and processed. Under this model, recommendation systems evolve from correlation-based engines toward reasoning-oriented architectures, supported by richer, more timely data flows that agents can both consume and help generate. In parallel, spending architectures are expected to tilt toward machine consumption models, privileging usage-based data infrastructure, observability, and real-time control engines over traditional seat-based metrics.
As agents assume more operational tasks, Meta argues that autonomy must be balanced with governance rather than treated as an unbounded capability. The company introduces “trusted data environments” in which internal agents can explore enterprise data while every output is traceable back to its source for scrutiny. Sensitive fields are masked before agents gain access, and each request is evaluated in real time against declared purposes and authorization rules. This approach aims to keep human oversight embedded in workflows even as machine-driven decision making expands.
The early impact of these changes is already visible inside Meta’s analytics culture. Within three months of launch, nearly two thirds of new internal dashboards were created through agentic data applications, reshaping how employees interact with information. For enterprises broadly, infrastructure readiness is framed not as a minor optimization but as a prerequisite for capturing value from AI agents before bottlenecks solidify into reality.
