Meta warns: 20 months to adapt infrastructure for AI agents

What happened?

At the VB Transform 2026 conference, Barak Yagour, Meta's VP of Infrastructure Engineering, issued a stark warning: organizations have roughly 20 months to transform their enterprise infrastructure, originally designed for humans, to support the growing demand for AI agents. Yagour revealed that agent queries in Meta's data systems multiplied by 30 in just six months—a reversal that breaks the assumptions on which the company built its systems over two decades.

The phenomenon is not unique to Meta. According to Imperva's 2025 report, automated traffic surpassed human traffic on the internet last year, reaching 51% of the total. Additionally, HUMAN Security's 2026 report notes that this traffic grows eight times faster than human traffic. Yagour cited both figures to describe an inflection point already underway. This shift is not gradual; it's an explosion. For perspective, in 2020, automated traffic accounted for less than 30% of total traffic, according to Akamai studies. Today, malicious and benign bots dominate the web, and the arrival of generative AI agents further accelerates the trend.

Why is it important?

Current infrastructure was built under assumptions that no longer hold. Yagour identified three critical areas collapsing simultaneously: capacity, identity, and speed. These three pillars, which supported systems like relational databases, CI/CD pipelines, and access control systems, are now faltering because they were designed for human users with predictable behavior patterns.

Capacity

Previously, one engineer generated one unit of load. Now, that same engineer can create 10 agents, each with its sub-agents, multiplying the load by 100. An organization of 1,000 people can generate the traffic of 100,000 users overnight. The solution is not to block agent traffic but to make infrastructure agent-aware, with dynamic controls that understand hierarchies, cost attribution, and priorities. Meta already faces query spikes 30 times higher than six months ago, forcing a rethink of horizontal scaling and load balancing. In comparison, the microservices boom of the 2010s multiplied internal traffic between services, but never at this exponential rate.

Identity

Agents don't fit traditional access control categories. They are not human, don't carry credentials, and are not deployed services, yet they make autonomous decisions. This demands a rethink of identity management. Yagour pointed out that current IAM systems (like Active Directory or Okta) cannot handle non-human identities acting on behalf of multiple users. New models like “agentive identities” with dynamic permissions and action auditing are needed. For example, an agent accessing a customer database must be traceable back to an employee's original request, but also have context restrictions. This problem is similar to the one that arose with cloud service accounts, but amplified by agent autonomy.

Speed

GitHub Copilot already writes 46% of the average code, but the rest of the pipeline (build, test, deploy, monitor) doesn't speed up just because a machine is the author. Code generation speed doesn't translate to delivery speed if infrastructure doesn't adapt. Yagour mentioned that traditional CI/CD pipelines, designed for human commits, become saturated when agents submit hundreds of changes per minute. This forces test parallelization, smart caching, and continuous deployments with automatic approval for low-risk changes. In 2024, GitHub reported that developers using Copilot complete tasks 55% faster, but the bottleneck is now in delivery infrastructure.

Consequences and necessary actions

Meta is already taking action. In February 2026, it launched agentive data applications, and within three months, 63% of published dashboards in the company were built with those tools. This raises a governance problem: human analysts acted as quality filters between data and decisions; now agents gain independence, requiring trusted data environments and oversight. Yagour emphasized that infrastructure must be “agent-aware”: capable of tracking consumption back to the originating use case, applying adaptive throttling, and ensuring agents operate within safety barriers. The 20-month window is a call to action for companies to start redesigning their systems before pressure becomes unsustainable.

The business consequences are profound. Companies that don't adapt could face runaway infrastructure costs (e.g., cloud bills doubling without warning), security breaches from poorly managed identities, and software delivery delays. Conversely, those investing in agent-aware infrastructure will gain competitive advantages: they can scale operations with smaller teams, launch products faster, and deliver personalized experiences at scale. Sectors like fintech, healthcare, and logistics, where precision and regulatory compliance are critical, must prioritize agent governance.

What should readers know?

• AI agent traffic is growing exponentially and already dominates the internet: from 30% in 2020 to 51% in 2025, and accelerating.
• Current systems (databases, CI/CD, access control) are not designed for autonomous agents, causing bottlenecks in capacity, identity, and speed.
• Companies must invest in agent-aware infrastructure with capacity, identity, and speed controls, including cost attribution, dynamic identities, and parallelized pipelines.
• Data governance and human oversight will remain critical but must evolve toward exception-based supervision, where humans review only high-risk decisions.
• The estimated 20-month timeline is an opportunity to act before the gap becomes critical. Companies that start now can avoid disruptions and capitalize on the agentive wave.

In summary, enterprise infrastructure is at a historic crossroads. Yagour's warning is not just for Meta: it's an alert for all organizations that depend on technology. Ignoring the agentive transformation could lead to a silent collapse of systems, while decisive action will open the door to a new era of efficiency and automation.