Meta Warns AI Agents Infrastructure May Crack in 20 Months

Enterprise AI agents were supposed to make work faster. Meta is warning that they may first make enterprise infrastructure break.

That is the real signal from Barak Yagour, Meta VP of Engineering for Data Infrastructure, who told VB Transform 2026 that systems built around human users are being strained by machine actors that query, code and consume data at a different tempo, according to VentureBeat. The primary problem is no longer whether AI agents can do useful work. It is whether AI agents infrastructure can survive when agents become a mass user population.

Yagour opened wearing Ray-Ban Meta AI glasses, a neat visual cue that AI has moved off the screen and into physical life. But his sharper point was buried under the hardware symbolism: enterprise systems still assume people are the main consumers of compute, data and permissions.

"What happens to the infrastructure we've spent years building when agents and not humans become the main consumers of that," Yagour said. "That's the world we're stepping into."

Meta's 20-month AI agents infrastructure warning turns assistants into load generators

Yagour’s closing line was blunt: Meta and the wider industry spent 20 years building infrastructure for humans, and may have 20 months to rebuild it for humans and agents working together at scale.

That timeline matters because the agent shift changes the basic unit of demand. A human employee used to map roughly to a predictable workload. An engineer generated requests, wrote code, queried systems and consumed data at human speed. Agents compress that cycle.

The expectation was simple: agents help employees do more. The reality Yagour described is harsher: agents multiply the number of system actors that infrastructure must identify, govern, throttle and charge back to a business purpose.

A cleaner way to frame the shift:

• Before: Human users drove most access patterns, capacity planning and governance checks.
• After: Agents can create more agentic workload than the humans who initiated them.
• Before: Identity meant a person, a role or a deployed service.
• After: An agent may not fit any of those categories cleanly.
• Before: Faster coding mostly meant faster developers.
• After: faster code generation can jam build, test, deploy and monitoring pipelines.

XOOMAR analysis: this is why the phrase AI agents infrastructure is becoming more useful than “AI tooling.” The hard part is not the chat interface. It is the plumbing behind it.

The 30x query spike shows why capacity planning is cracking

The most important number in Yagour’s talk was 30x. He said agentic queries hitting Meta’s data systems grew 30x in a single half, a rate that breaks the tidy assumptions behind normal capacity forecasts.

The wider traffic data points in the same direction. Yagour cited Imperva's 2025 Bad Bot Report, which found automated traffic overtook human traffic on the internet last year, reaching 51% of the total. He also cited HUMAN Security's 2026 State of AI Traffic report, which said that traffic is growing roughly eight times faster than human traffic.

Inside a company, the multiplier can look even stranger.

"One engineer used to mean one unit of load," Yagour said. "Now one engineer spawns 10 agents, each spawning subagents. Your 1,000-person org can generate the load of 100,000 users practically overnight."

That line explains why Meta is not talking about blocking agents. It is talking about making infrastructure agent-aware.

Yagour’s proposed answer includes:

• Hierarchy tracking: Systems need to understand which agent spawned which subagent.
• Cost attribution: Consumption must trace back to the use case that created it.
• Dynamic throttling: Controls should adjust based on priority rather than treat all requests alike.
• Agent-aware capacity controls: Infrastructure has to recognize non-human demand patterns.

Meta’s own targets show the scale of the data problem. Yagour said the company is building toward 500 million queries per second and a petabyte per second of throughput for training data reads. Most companies do not operate at Meta’s scale, but the bottleneck category is the same: agents make hidden infrastructure assumptions visible.

For a related enterprise deployment lens, see XOOMAR’s coverage of how Agent Orchestration Runs Ahead of Real Enterprise AI.

Capacity, identity and CI/CD now fail as one system

Yagour grouped the breakage into three buckets: capacity, identity and velocity. The useful part is that he did not treat them as separate IT concerns. In agentic systems, they compound.

Identity may be the most awkward of the three. Yagour said an agent is not a human user, does not carry a badge and is not a deployed service, yet it makes decisions on its own. That makes traditional access rules feel misaligned.

Velocity has the same mismatch. Yagour cited a company-reported figure that GitHub Copilot writes 46% of the average user's code, then pointed out that the rest of the software delivery chain does not magically accelerate.

"That code still needs to be built, tested, deployed, monitored," he said. "The agent writes the code in seconds, but your CI/CD pipeline doesn't get faster just because the machine is the author."

XOOMAR analysis: the productivity story around coding agents is incomplete if the delivery system remains human-paced. Code generation is only one lane. The merge, test, deploy and monitoring lanes still decide whether speed turns into output or backlog.

This also connects to vendor pressure around enterprise AI packaging, a theme we covered in Microsoft AI Models Turn on OpenAI in Risky Sales Push.

Trusted data environments make governance part of the data plane

Data is where Yagour’s warning becomes operational. He said "Data sits at the center of everything," because it feeds decisions, products, recommender systems and next generation models.

Meta’s own adoption curve shows the pull. In February, the company shipped what Yagour called agentic data apps. Within three months, 63% of dashboards published across Meta were built using the new tooling.

That is a rapid shift from human-mediated analysis toward agent-assisted data work. It also creates the chaos risk Yagour named directly.

"Autonomy without governance is nothing but chaos," he said.

Meta’s answer is trusted data environments. Inside them, agents can explore data, but sensitive fields are masked before access, every request is evaluated in real time and outputs are traced back to source material.

Yagour’s operating principle is simple: explore broadly, release narrowly.

XOOMAR analysis: that makes governance part of infrastructure, not a review step bolted on after analysis is complete. If agents are going to work inside data systems, the control layer has to sit where the work happens.

Reasoning models are forcing Meta past batch ETL

Meta’s data layer is also changing because its models are shifting from correlation toward reasoning. Yagour put it plainly:

"Reasoning is data hungry."

Pattern-matching systems can work from sparse and summarized signals. Reasoning models, as Yagour described them, need fuller behavioral history: interactions across surfaces and over time, plus current intent.

That creates two infrastructure moves at Meta.

First, real-time streaming is replacing batch ETL for ranking pipelines. Yagour said a 24-hour pipeline does not work when a model is reasoning about what a user wants now. Real-time streaming is becoming the backbone of Meta’s ranking and recommendation systems.

Second, storage is becoming schema-aware. Yagour said Meta previously stored user data as opaque blobs, which caused heavy overfetching and idle GPU capacity. The new direction is storage that knows what it holds, so systems can pull only the columns and time ranges a query needs.

This feeds Meta’s recommendation ambitions. Yagour said 42% of Instagram users have told the company they want to fundamentally change the algorithm, not just tweak one setting or session. Meta’s response is what he called fully conversational recommendations, where users state what they want more of and the system reasons about intent.

His soccer example makes the point: the same search term could return different results for a casual fan looking for highlights and a club athlete seeking training drills.

The rebuild hits engineers, security teams, CFOs and Instagram users differently

The agent rebuild does not land evenly.

For infrastructure leaders, the problem is capacity shock. A small number of humans can now create large volumes of machine demand.

For security teams, the problem is identity. Agents are decision-making actors that do not fit cleanly into old access models.

For finance teams, XOOMAR analysis: Yagour’s emphasis on cost attribution points to a looming spending-control issue. If agent consumption cannot be traced to the use case that spawned it, internal budgeting gets blurry fast.

For developers, the promise is speed, but the friction sits in CI/CD. Code may arrive faster than systems can safely process it.

For users, Meta’s conversational recommendations could offer more direct input into feeds. The unresolved question is how much intent a platform should infer, how clearly it explains that inference and how tightly users can correct it.

Yagour described agents, data and recommendations as linked rather than separate.

"Agents make data more accessible. Better data makes reasoning. Reasoning creates new demands that push agents and infrastructure forward," he said. "This isn't linear; it's a flywheel."

The 20-month clock favors agent-ready plumbing over demos

Yagour said Meta is experimenting at every level, including whether SQL is the right interface for agents at all. He also said Meta’s storage already operates in the multi-digit exabyte range and needs to keep expanding.

That does not mean every company should copy Meta’s architecture. It does mean the questions Meta is asking are likely to become standard for serious agent deployments.

The evidence to watch is practical:

• Agent identity: Can companies define what an agent is, what it can access and who is accountable for it?
• Agent budgets: Can they attribute cost to the use case that created demand?
• Governed data zones: Can agents explore without exposing sensitive fields or untrusted outputs?
• Priority throttling: Can infrastructure distinguish critical agent work from low-value load?
• Interface shifts: Do agents keep using SQL, or do new data interfaces emerge for machine reasoning?

Yagour’s warning is not that agents are useless. It is that useful agents will stress systems built for a different era. The firms to watch are not the ones with the flashiest demos. They are the ones rebuilding AI agents infrastructure fast enough to let humans and agents work together without turning autonomy into chaos.

Impact Analysis

• AI agents could overwhelm enterprise systems designed around slower human usage patterns.
• Companies may need to rethink identity, governance, throttling and cost controls for machine actors.
• Meta’s warning suggests the infrastructure shift could happen far faster than past enterprise technology transitions.