Meta executive: Companies have 20 months to adapt infrastructure for AI agents

Meta Platforms' VP of Engineering, Barak Yagour, estimates that companies have about 20 months to retool their infrastructure for the demands of AI agents. Speaking at the VB Transform 2026 conference, Yagour argued that current enterprise infrastructure, built for human users, is showing strain under the load of autonomous agents.

Yagour reported that agentic queries hitting Meta's data systems surged 30-fold in a single half-year. This dramatic increase challenges two decades of assumptions about infrastructure design. This trend is not isolated to Meta; automated traffic surpassed human traffic online last year and is growing roughly eight times faster, according to industry reports.

The shift is simultaneously breaking three core infrastructure assumptions at Meta: capacity, identity, and velocity. Yagour explained that where one engineer previously represented one unit of load, a single engineer can now spawn multiple agents, leading to exponential increases in demand. Furthermore, AI agents do not fit traditional user identity models, as they are neither human users nor deployed services.

Meta's strategy is not to block AI traffic but to make its infrastructure "agent-aware." This involves implementing dynamic controls that understand agent hierarchies and attribute costs to specific use cases. The company is also developing "trusted data environments" to allow agents more autonomous data exploration while maintaining governance and security, including masked sensitive fields. Meta is also evolving its data layer to be more real-time and schema-aware to better support the data requirements of AI models for decision-making and recommendation systems.