Companies have embraced the agentic AI boom, to the point that some They are drowning in so many AI agents. Employees are creating uncontrolled agents, triggering token consumption and causing many of those agents to duplicate tasks. But the real underlying problem is something else: the infrastructure on which all this is running is not prepared to support it.
Meta’s warning. They tell it in Venture Beat. During the VB Transform 2026 talk, Meta’s vice president of engineering Barak Yagour gave a very powerful figure: queries from agents reaching the Meta system have multiplied by 30 in the last six months. “What happens to the infrastructure we’ve built over years when agents, not humans, become its primary consumers?” Yagour asked the audience. The answer is even more worrying: “We have spent 20 years building infrastructure for humans. Maybe we have 20 months to rebuild everything for a world where humans and agents co-create at scale.” And he adds that “The opportunity is open, but it will not last long.”
lthe three problems. The emergence of the agents threatens to break the infrastructure on three fronts:
- Ability: Until recently, to plan how much computing power, storage or bandwidth was needed to carry out a project, it was enough to count the number of employees working on it. Yagour explains it very graphically: “Now, an engineer generates 10 agents, and each one generates sub-agents. An organization of 1,000 people can generate the load of 100,000 users practically overnight.”
- Identity: Access control systems are not designed for an AI agent. It is not human, but it moves around the infrastructure and makes decisions on its own.
- Speed: The third problem is the speed at which the agents move. They can write code faster than any human, but “that code still needs to be compiled, tested, deployed and monitored.”
Deep changes. Yagour mentions that reasoning models are changing the way the data layer is processed. Detecting matches from keywords is not the same as reasoning about someone’s intention; That requires a complete history of behavior, not a summary. This requires making changes to how that data is processed and stored.
Meta is moving from batch processing, which can take up to 24 hours to update, to real-time processing as it is key when the model is reasoning about what a user wants at that very moment. On the other hand, they are moving from an opaque storage to one that understands its own content, so that it only reads what it needs in each query and thus lightens the load on the GPUs.
Setting limits. According to Yagour “Autonomy without governance is nothing more than chaos”, which is why Meta’s solution is to create what they call “trusted data environments.” Here, agents can move freely “but every result is traced back to its source and carefully vetted. This way you can always be sure that the data shared is trusted and controlled.” It is a way of giving freedom to the agents, but in a controlled way and without exposing themselves to the risk of ending up messing up.
Image | Xataka with Magnific
In Xataka | AI agents can destroy jobs. Or they can also save a century-old cheese factory from bankruptcy
GIPHY App Key not set. Please check settings