- In a VentureBeat survey of 166 organizations, only 2% of orchestration platform users chose a platform mainly because of the AI model behind it.
- 67% expect agent control to sit at least partly outside a provider-managed service. 23% of orchestration users track agent spending only after the fact.
- Leaders should ask who controls agent permissions, logs and spending limits, and what it would cost to move agents off their primary platform.
The model is not what buyers say they are buying
OpenAI, Anthropic and Google now sell platforms for building and running AI agents, next to their models. The obvious guess is that companies pick the platform that matches the model they like. A new survey says that is rarely the stated reason.
VentureBeat Intelligence ran its August VB Pulse survey with respondents from organizations of 100 or more employees. It asked 162 platform users to name the single factor that mattered most. Just 2% named the pull of a leading AI model, such as choosing Anthropic's platform in order to use Claude.
The winning answers were all practical. Support for many models and tools drew 22%. Stable performance in production drew 20%. Simpler building drew 19%. Tighter say over how agents run drew 18%.
The lesson here is that buyers describe choosing the platform around the model: its reach, reliability and ease of use, plus some control. None of those is the model. Each platform choice also raises governance questions about who can do what, and who can stop it.
What an orchestration platform actually does
Think of an orchestration platform as a dispatcher for software agents. It assigns work, connects agents to tools and keeps each one within the access it was given.
The control plane is the rulebook and switchboard on top of that. It sets which agent acts, what data and tools it may touch, and what record is kept.
A control plane run by the model provider is easier to start with, but it binds your rules to one vendor. One you own can work across vendors. That freedom has a price: someone inside the company must acquire or construct that layer and operate it.
The survey asked where respondents expect that control to sit by the end of 2026. Two in three (67%) pointed to a hybrid, an external platform or a custom in-house build. Only 27% expect a provider-managed service. No answer won a majority, and hybrid led at 33%.
Model makers still win real usage
This is not a story of model makers losing. OpenAI's platform has the widest reach in the sample: 45% of all respondents use its Agents SDK or Responses API, and 69% of those users call it their primary platform. Google's Enterprise Agent Platform is used by 39%, and 57% of its users rely on it first.
Anthropic's platform more often plays second fiddle. Claude Platform appears in 27% of respondents' stacks, but only 38% of its users treat it as their primary one. The other 62% lean on a different platform first.
Looking ahead, 100 respondents plan to add or swap a platform within 12 months. Of those, 40% named OpenAI's platform and 35% named Anthropic's Claude Agent SDK or Managed Agents. These planning figures cannot be compared with the usage figures above. The usage question listed Claude Platform but not the Agent SDK.
Many firms run several platforms, and the average respondent runs 1.73. Firms with two or more are likelier to plan another change: 84%, against 46% of single-platform firms. Every extra platform adds its own access rules, monitoring and billing. Agents on one platform may not see what agents on another are doing.
The cost question is still open
Picture an agent stuck in a retry loop, or passing work back and forth with other agents. Each step consumes tokens, so the bill grows with every pass. Almost one in four orchestration users (23%) review agent costs only once the money is spent. Those firms cannot halt an agent mid-run when it passes its budget, at least not from within their orchestration setup.
For these firms, a runaway agent can be caught only once the money is spent. The survey does not say who spots the overrun or how. The remaining 77% rely on built-in platform limits, their own gateways, or routing work to cheaper models.
Read the numbers with care
This is a survey of 166 respondents. It measures stated reasons and expectations, not what companies end up doing.
One result needs special caution. In July, 13% of platform users reported running one platform. In August, 59% did. The researchers call the gap statistically significant. But different people answered each month, so they cannot say whether enterprises changed.
True multi-agent systems look scarce in this sample. 53% of respondents say a quarter or fewer of their deployed agents fit that description. That group includes 10% who say all of theirs are essentially prompt wrappers. The researchers note that a single agent using tools may still do more than a chatbot wrapper.
Questions to put to your team
Which platform is our primary one, and what would moving our agents off it cost? Moving means rebuilding workflows and controls.
Where do agent permissions and logs live? If we run more than one platform, who sees across all of them?
Can we stop an agent in the middle of a run when it hits its budget, or do we find out afterward?
In this survey, the model was the reason almost nobody gave. The reasons buyers did give describe the platform around the model. That platform comes with governance questions your own team has to answer.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: VentureBeat.
