Frontier AI companies increasingly cite agent counts in the tens of thousands, and climbing. The number itself is presented as evidence of progress.
I sometimes wonder whether it belongs in a launch announcement or a risk register. What matters is what that scale creates and whether the enterprise can operate it.
Consider 1,000 agents across service management, finance, sales, HR, legal, and operations.
They will not arrive through one program. Many will come embedded in the CRM, ERP, ITSM, HCM, and contract platforms each function already runs. They will be useful locally and adopted independently, often with different prompts, memory, credentials, and policies.
Each function sees its own agents. No one sees or owns the system they form together.
The same applies to the runtimes beneath them. Each may look controlled inside its platform. But every additional runtime creates more surface area where enterprise IP has to be protected.
An individual agent may work well. The risk sits in the dependencies agents and runtimes create together, and in whether the enterprise can operate that collective system safely and economically.
Without shared foundations, context gets recreated and knowledge begins to diverge. Policies drift, permissions accumulate, and lineage and cost become harder to see.
A thousand agents could become a new form of enterprise sprawl. This time, the components reason, remember, and act. The layer also keeps changing as models, data, permissions, and ownership evolve.
At tens of thousands, sprawl can become the default architecture.
Four capabilities look structural to me:
1. A deliberate runtime estate with common control across it.
2. Shared context and knowledge, including governed memory.
3. An operating model for identity, ownership, lifecycle, and earned autonomy.
4. AgentOps across lineage, behavior, outcomes, drift, and cost.
Vendors can provide pieces. They cannot decide which runtimes to support, how context should be shared, or where accountability sits. Those choices have to reflect the enterprise's decision rights and risk boundaries.
Experiment broadly in controlled environments. That may mean fewer strategic runtimes, clearer ownership, tighter control boundaries, and a smaller portfolio that earns autonomy over time.
I am still working through what this takes in practice. No single function naturally owns the whole problem. And each runtime provider is incentivized to make its own platform work, not the estate around it. That may be exactly why the fabric does not get built.
The ten-thousandth agent is not the hard part. The first thousand may already have created an unmanaged operating layer beneath the enterprise.
The agent count will make the headline. The operating fabric will determine whether that scale becomes leverage or sprawl.
First published on LinkedIn, July 12, 2026.