Meta's infrastructure lead says there are maybe 20 months to rebuild for agents

Treat the timeline as one executive's estimate, and the observation underneath it as the part worth planning around.

Meta's infrastructure VP told VB Transform 2026 that companies have perhaps 20 months to rebuild their systems for agents. He cited a 30-fold rise in agentic queries over six months.

The quote travelled fast, mostly stripped of the detail that an infrastructure executive said it at a conference about infrastructure.

The 20-month figure is a personal estimate offered in conference remarks, so it's neither research nor modelled. The person saying it has a professional interest in infrastructure urgency.

The 30-fold growth in agentic queries is the more substantial claim, because Meta presents it as traffic it observes directly. It remains unaudited, stated by an executive at his own company's session with no published methodology.

The underlying argument is about assumptions. Systems built on the idea that a human reads the output behave badly when a machine consumes it at volume.

Discount the deadline, since nobody knows whether it's 20 months and the person saying so isn't neutral. The observation underneath it doesn't depend on the timeline being right.

At our clients' scale the practical version is unglamorous, because this is not about rebuilding your architecture at all. It's about whether your revenue data exists anywhere an agent could read it. Say prospect research lives in one person's notes and reply history sits in an inbox. No agent helps you then, however good agents get.

The companies that benefit are not the earliest adopters but the ones whose data was already connected, since that's the precondition. Everyone else spends the first three months of any agent project doing plumbing they had two years to do.

The companies that benefit are not the earliest adopters but the ones whose data was already connected.

Do the boring version now, since it pays off whether or not the prediction holds. Get prospect data, reply history and account context into systems with an API and a stable schema. It is useful immediately, with no agents involved.

Here's a reasonable check. Give a competent new hire read access to your tools and nothing else, then see whether they could reconstruct why you contacted a given prospect. If they can then an agent can, and if the answer lives in somebody's head you have a data problem wearing an AI costume.

None of this needs a rebuild or a deadline, because it needs the same wiring you wanted anyway. You are only doing it sooner than planned.

Working out how much of that plumbing you already have is what the Diagnostic maps.

Sources

How we choose and label these: our sourcing standard.