Meta VP: We have about 20 months to rebuild infrastructure for AI agents

Meta's VP of Engineering, Barak Yagour, warned that organizations have approximately 20 months to fundamentally retool their infrastructure to accommodate the rise of agentic artificial intelligence. Speaking at VB Transform 2026, Yagour highlighted that current enterprise infrastructure was designed for human users, not AI agents, and is rapidly becoming inadequate. He revealed that agentic queries on Meta's systems surged 30 times in just six months, disrupting decades-old assumptions. This trend mirrors broader internet shifts, with automated traffic surpassing human traffic last year and growing significantly faster.

Yagour identified three core infrastructure assumptions that are breaking: capacity, identity, and velocity. He explained that one engineer can now spawn numerous agents, exponentially increasing load beyond traditional calculations. Agentic identity also challenges existing access control systems, as agents are neither human users nor deployed services. Furthermore, while AI tools like GitHub Copilot accelerate code generation, they do not inherently speed up the entire development pipeline, creating bottlenecks.

To address these challenges, Meta is developing agent-aware infrastructure with dynamic controls, cost attribution, and adaptive throttling. The company is also focusing on 'trusted data environments' to govern AI agent access to data, masking sensitive fields and scrutinizing every output. Reasoning models are driving a shift from batch processing to real-time streaming for ranking systems and demanding schema-aware storage to prevent GPU starvation. Yagour concluded that this transformation is a reinforcing flywheel, with agents making data more accessible, better data enabling reasoning, and new demands pushing agents and infrastructure forward.