For the August VB Pulse, VentureBeat Intelligence surveyed respondents from organizations with 100 or more employees. We asked which agent orchestration platforms they run and why, what they plan to adopt next, where they expect agent control to sit and how they stop runaway agent spending.
OpenAI, Anthropic and Google now sell their own platforms for building and running AI agents, alongside their AI models. Few enterprises chose a platform because of the model behind it. Among enterprises that use a listed agent orchestration platform, only 2% said alignment with a leading AI model most influenced their choice of platform. Flexibility across models and tools, production reliability, ease of development and control over agent execution together drove 78% of those choices.
Yet the model makers' platforms are still widely used. Of all respondents, 45% use OpenAI's Agents SDK or Responses API, and 69% of those OpenAI users named OpenAI's platform primary. Of the enterprises running Anthropic's Claude Platform, 38% named it their primary orchestration platform, and 62% named another platform.
Many enterprises already run more than one platform, and those enterprises are far more likely to add or replace one. Among enterprises running two or more platforms, 84% plan a new, additional or replacement platform within 12 months, against 46% of single-platform enterprises.
In this August sample, 59% of platform users run one platform. In our July Pulse, only 13% of platform users reported one platform. The gap between July's 13% and August's 59% is statistically significant, but different respondents answered in each month.
The control plane is the system that directs and governs enterprise agents. Of all respondents, 67% expect the control plane to sit at least partly outside a provider-managed service by the end of 2026. Only 27% expect a provider-managed agent service. Asked about the top risk if orchestration control lives inside a model provider's platform, 28% of respondents chose inflexibility across models and tools.
True multi-agent workflows are still a small share of the agents at many enterprises. Of all respondents, 53% said a quarter or fewer of their deployed agents are true multi-agent orchestrated workflows.
Of enterprises using agent orchestration, 23% track agent spending only after the fact, so they have no real-time way within their orchestration pipelines to stop an agent before it exceeds its budget.
Finding 1. Only 2% of enterprises chose their AI agent platform mainly for the AI model behind it
Flexibility across models and tools, production reliability, ease of development or control over agent execution most influenced 78% of platform choices; the model behind the platform most influenced 2%.
An agent orchestration platform is the software that coordinates a company's AI agents: it hands out tasks, calls tools and keeps agents within their permissions. OpenAI, Anthropic and Google each sell one alongside their own models. Independent platforms, such as LangChain, are built to work with models from many providers.
We asked the enterprises that use a listed platform which factor most influenced their orchestration platform selection.
Finding 1 — Only 2% of enterprises chose their AI agent platform mainly for the AI model behind it
Base: 162 enterprises that use a listed platform. One answer each.
Flexibility across models and tools was named by 22% of platform users, and production reliability by 20%. Ease of development was named by 19% of platform users, and control over agent execution by 18%.
Only 2% of platform users named "model gravity." In the questionnaire, we defined model gravity as native alignment with a state-of-the-art base model, such as choosing Anthropic's orchestration platform to use Claude.
Finding 2. 62% of enterprises using Anthropic's Claude Platform make another company's agent platform primary, so enterprises often run Anthropic's platform as a secondary platform
Of the enterprises that use OpenAI's Agents SDK or Responses API, 69% named OpenAI's platform primary.
The primary orchestration platform is the one an enterprise relies on first to build, run and govern its agents. Moving agents off a primary platform later means rebuilding their workflows and controls.
We asked which platforms each organization uses, then which one is its primary orchestration platform.
Finding 2 — 62% of enterprises using Anthropic’s Claude Platform make another company’s agent platform primary, so enterprises often run Anthropic’s platform as a secondary platform
Base: all 166 respondents, each of whom answered the platforms-in-use question; respondents could name several platforms. The users column uses each platform's users as its base. The primary-pick column uses the 154 enterprises that named a primary platform they also use. See Methodology for respondents whose primary platform was one they had not ticked. The remaining 4% of primary picks (Base: 154) went to LangChain / LangGraph, LlamaIndex, custom in-house orchestration and independent enterprise agent platforms.
Of all respondents, 45% said they use OpenAI's Agents SDK or Responses API, and 69% of OpenAI's users named it primary. Of all respondents, 39% use Google's Enterprise Agent Platform, and 57% of Google's users named it primary.
Anthropic's Claude Platform is used by 27% of respondents, but only 38% of Claude Platform users named it primary. Of all Claude Platform users, 27% named OpenAI's platform primary, 18% named Google's and 18% named a different platform.
When we leave out the respondents who named a primary platform they had not ticked as used, 41% of the remaining Claude Platform users named Anthropic's platform primary.
Microsoft's AI Foundry or Copilot Studio is used by 19% of respondents, and Salesforce's Agentforce or another enterprise app platform by 15%. Of the enterprises using Microsoft's platform, 68% named it primary, and so did 68% of those using Salesforce's.
Among the enterprises that named a primary platform they also use, 34% named OpenAI's and 24% named Google's. Microsoft's platform was named by 14%, and Anthropic's and Salesforce's by 11% each.
Finding 3. Multi-platform enterprises are far more likely to plan a new, additional or replacement orchestration platform
Among enterprises running two or more platforms, 84% plan a platform change within 12 months, against 46% of single-platform enterprises.
Running more than one orchestration platform gives an enterprise access to different tools and models, but each platform brings its own permissions, monitoring and billing. Teams then have to govern agents across several systems, and an agent on one platform may not see what an agent on another is doing. Each new or additional platform adds another system for security and operations teams to connect and watch.
We asked which agent orchestration platforms each organization uses today, allowing several answers.
Finding 3 — Multi-platform enterprises are far more likely to plan a new, additional or replacement orchestration platform
Base: 162 enterprises that named at least one listed orchestration platform they use today. The average is 1.73 platforms per enterprise.
Of the enterprises that use orchestration, 59% named one platform. Another 23% named two, and 19% named three or more.
In our July Pulse, by contrast, only 13% of platform users said they ran one platform. The gap between July's 13% and August's 59% is statistically significant. However, the two months drew different groups of respondents, so we cannot say whether enterprises changed.
We then asked whether each organization plans to adopt a new, additional or replacement orchestration platform in the next 12 months. Enterprises that already run two or more platforms were far more likely to say yes. Of multi-platform enterprises, 84% plan a change, against 46% of single-platform enterprises.
Finding 3 — Share planning a new, additional or replacement platform within 12 months
Base: 162 enterprises that use a listed platform, split by number of platforms. The July row uses July's 105 platform users. Each cell gives its own base.
Multi-platform enterprises also plan a change more often within each group we checked. Among final decision-makers, 92% of multi-platform enterprises plan a change, against 66% of single-platform enterprises.
Among respondents who are not final decision-makers, 71% of multi-platform enterprises plan a change, against 29% of single-platform enterprises.
Multi-platform enterprises also plan a change more often at both larger and smaller organizations. At organizations with 2,500 or more employees, 77% of multi-platform enterprises plan a change, against 28% of single-platform enterprises.
At organizations below 2,500 employees, 88% of multi-platform enterprises plan a change, against 55% of single-platform enterprises.
Among July's platform users, 71% of multi-platform enterprises planned a change, against 50% of single-platform enterprises. The July gap is too close to call.
Finding 4. 60% of enterprises plan a new, additional or replacement orchestration platform within a year, and 28% expect to adopt one within three months
Of all respondents, 60% plan to adopt a new, additional or replacement platform within 12 months, and 28% expect to do so within three months.
Adopting a new orchestration platform means moving or rebuilding agent workflows, connecting the platform to company data and tools, and setting up its permissions and monitoring. An additional platform also adds another system to govern. A replacement platform means moving workflows off the old platform without breaking agents already in production.
We asked whether each organization plans to adopt a new, additional or replacement orchestration platform in the next 12 months.
Finding 4 — 60% of enterprises plan a new, additional or replacement orchestration platform within a year, and 28% expect to adopt one within three months
Base: all 166 respondents. One answer each.
Of all respondents, 19% expect to adopt a new, additional or replacement platform within three to six months, and 13% within six to 12 months.
We then asked the enterprises planning a change which platforms they are considering.
Finding 4 — Platform under consideration
Base: 100 enterprises planning to adopt a new, additional or replacement platform within 12 months. Respondents could name several platforms. "Do not use it today" means the respondent did not tick that platform in the platforms-in-use question. We listed Anthropic's Claude Platform in that question, but not its Agent SDK.
Of the enterprises planning a change, 40% named OpenAI's platform. Of the enterprises considering OpenAI, 68% already use OpenAI's platform.
About half of the enterprises considering Anthropic do not use Anthropic's Claude Platform today. Of the enterprises planning a change, 35% are considering Anthropic's Claude Agent SDK or Managed Agents. Of the enterprises considering Anthropic, 54% do not use Claude Platform today, and 46% already do.
Finding 5. 67% of enterprises expect the agent control plane to sit at least partly outside a provider-managed service by the end of 2026
Of all respondents, 67% chose a hybrid, external or custom in-house control plane, against 27% who chose a provider-managed agent service.
The agent control plane decides which agent runs, which tools and data each agent can reach, and what gets logged. A provider-managed control plane runs inside a provider's own agent service, which can be simpler to set up but ties agent governance to that provider. An external or in-house control plane can work across models from several providers, but the enterprise has to build or buy it and keep it running.
We asked where respondents expect the primary control plane for enterprise agents to live by the end of 2026.
Finding 5 — 67% of enterprises expect the agent control plane to sit at least partly outside a provider-managed service by the end of 2026
Base: all 166 respondents. One answer each.
No single answer was chosen by a majority of respondents. A hybrid approach, combining a model provider's own agent tooling with external orchestration, was chosen by 33%. A provider-managed agent service was chosen by 27%.
Taken together, 67% of respondents chose a hybrid approach, an external orchestration platform or a custom in-house control plane.
We also asked which risk concerns each organization most if orchestration control lives inside a model provider's platform.
Finding 5 — Top risk if orchestration control lives inside a model provider’s platform
Base: all 166 respondents. One answer each.
Of all respondents, 28% chose "inflexibility across models and tools." "Security and permissioning limitations" and "limited visibility and observability" were each chosen by 22%.
Only 16% of respondents chose vendor lock-in. And 13% answered "This is not a concern for us."
Finding 6. 53% of enterprises say a quarter or fewer of their deployed agents are true multi-agent workflows, and 81% say half or fewer
Of all respondents, 53% put a quarter or fewer of their deployed agents in true multi-agent workflows, and 10% say none are.
A true multi-agent workflow splits a task among several agents that hand work to each other, with an orchestration layer coordinating them. A chatbot wrapper sends one prompt to one model and returns the answer. Multi-agent workflows can take on longer business processes, but they add more places for an error to spread and more model calls to pay for.
We asked respondents what share of their organization's deployed "agents" are true multi-agent orchestrated workflows. In the question, we contrasted true multi-agent workflows with "simple, single-prompt chatbot wrappers."
Finding 6 — 53% of enterprises say a quarter or fewer of their deployed agents are true multi-agent workflows, and 81% say half or fewer
Base: all 166 respondents. One answer each.
A single agent that uses tools or runs a multi-step workflow is neither a "true multi-agent orchestrated workflow" nor a "simple, single-prompt chatbot wrapper," the two descriptions in the question. An agent outside the multi-agent share may therefore still do more than a chatbot wrapper.
Of all respondents, 53% said a quarter or fewer of their agents are true multi-agent workflows. The 53% includes the 10% who said every one of their deployments is essentially a chatbot or prompt wrapper. At the other end, 19% said more than half of their agents are true multi-agent workflows.
Finding 7. 23% of enterprises using agent orchestration track agent spending only after the fact, with no real-time way within their orchestration pipelines to stop a runaway agent
Of enterprises using agent orchestration, 77% chose native platform controls, custom gateways or low-cost model routing instead of monitoring only after the fact.
An agent that loops, retries or calls other agents can run up model charges quickly, because every step uses tokens. A real-time spending control stops an agent when it reaches a budget limit. Monitoring after the fact shows what an agent spent, but only after the charges have been incurred.
We asked the enterprises that use a listed platform how each uses its orchestration layer to control spending on tokens.
Finding 7 — 23% of enterprises using agent orchestration track agent spending only after the fact, with no real-time way within their orchestration pipelines to stop a runaway agent
Base: 162 enterprises that use a listed platform. One answer each, so respondents who chose another answer may also monitor after the fact. Respondents who do not use agent orchestration are left out.
Of platform users, 23% said they track spending only through logs after the fact. Respondents who monitor only after the fact have no real-time, programmatic way within their orchestration pipelines to stop an agent before it exceeds its budget.
Of platform users, 77% chose one of three other approaches. Twenty-eight percent rely entirely on the budget caps and limits built into their primary orchestration platform.
Another 28% said their orchestration framework lacks granular financial controls, so they built their own gateways or middleware to stop runaway agent loops. And 20% said they route token-heavy work to low-cost models.
Finding 8. 34% of enterprises name agent workflow tooling as the orchestration investment that will grow most next year, and only 5% say their budget is not increasing
Of all respondents, 34% named agent workflow tooling as the orchestration investment that will grow most next year, and 25% named security and permissions enforcement.
Agent workflow tooling is the software teams use to design, build and connect the steps an agent follows. Security and permissions enforcement controls what each agent can reach, and monitoring and debugging show what agents did and why they failed. Each area covers a different stage of running agents, from building them to keeping them in bounds to fixing them.
We asked which orchestration-related investment "will grow most next year."
Finding 8 — 34% of enterprises name agent workflow tooling as the orchestration investment that will grow most next year, and only 5% say their budget is not increasing
Base: all 166 respondents. One answer each.
Of all respondents, 34% named agent workflow tooling. Security and permissions enforcement was named by 25%, and agent monitoring and debugging by 20%. Only 5% of respondents said their budget is not increasing.
What changed since the July Pulse
We asked the same questions with the same answer options in July and August, apart from the primary-platform question. The two months drew different groups of respondents, so we cannot say whether the same enterprises changed.
Technology and software companies made up 53% of July respondents and 10% of August respondents. Organizations with 2,500 or more employees made up 70% of July respondents and 34% of August respondents. Final decision-makers for AI purchases made up 27% of July respondents and 50% of August respondents.
What changed since the July Pulse — Who answered
Base: all qualified respondents in each month (107 in July, 166 in August).
What changed enough to call
What changed since the July Pulse — What changed enough to call
Enough to call: on each measure, the gap between July's and August's respondents is larger than chance alone would likely produce. Rows on platform users use bases of 105 in July and 162 in August.
Rows on planners use respondents planning a change: 72, then 100. The other rows use all respondents: 107, then 166.
What did not change enough to call
What changed since the July Pulse — What did not change enough to call
Too close to call: chance alone could produce each gap. The two planner rows use respondents planning a change: 72 in July, then 100.
The selection-factor row uses platform users: 105, then 162. The first row uses all respondents: 107, then 166.
What we did not compare
Primary-platform shares. We accepted several answers to the primary-platform question in July and one in August, so the two months measured different things.
Spending control after the fact. In July, 21% of all respondents said they monitor spending only after the fact, and July's base included respondents who use no listed platform. August's figure counts platform users only, so the two bases differ.
Primary success metric. We do not carry July's counts for the success-metric question in this report, so we did not test that question against July.
Satisfaction ratings. In August, 25% of platform users skipped the rating questions, and 80% of the platform users who skipped run a single platform. No July respondent skipped them, and we do not use the August ratings.
The bottom line: Enterprises choose AI agent platforms for flexibility, reliability, ease of development and control, and only 2% choose one mainly for the model behind it
Only 2% of platform users named the model behind their platform as the main reason they chose it. Of OpenAI's platform users, 69% name OpenAI's platform primary, and 38% of the enterprises using Anthropic's Claude Platform name it primary.
Most platform users in this survey run one agent orchestration platform. But 56% of the enterprises planning a new, additional or replacement platform already run two or more platforms.
Most respondents expect at least part of the agent control plane to sit outside a provider-managed service by the end of 2026. And 23% of enterprises using agent orchestration track agent spending only after the fact.
In the next Pulse, we will retest whether multi-platform enterprises are far more likely than single-platform enterprises to plan a platform change. We will also list Anthropic's products under the same names in both platform questions, so current and prospective Anthropic users can be counted on the same terms.
Respondent profile
Respondents | Share (of 166) |
Company size: 100 to 499 employees | 39% (65) |
Company size: 500 to 2,499 employees | 27% (44) |
Company size: 2,500 to 9,999 employees | 19% (32) |
Company size: 10,000 to 49,999 employees | 5% (9) |
Company size: 50,000 or more employees | 10% (16) |
Purchasing role: final decision-maker for AI | 50% (83) |
Purchasing role: recommender or influencer | 32% (53) |
Purchasing role: user | 13% (22) |
Purchasing role: no involvement | 5% (8) |
Industry: healthcare and life sciences | 16% (26) |
Industry: manufacturing and industrial | 15% (25) |
Industry: retail and consumer | 14% (24) |
Industry: technology and software | 10% (17) |
Industry: financial services | 8% (13) |
Industry: all others | 37% (61) |
Job title: consultant or advisor | 8% (14) |
Base: all 166 respondents. Company size, purchasing role and industry each add up to 166. All other industries: education 7%, professional services 6%, government and public sector 6%, media and telecom 1%, and "Other" 17%.
Methodology
VentureBeat Intelligence fielded this VB Pulse survey in August to a self-selected audience of VentureBeat readers, so the results do not come from a probability sample. We received 221 responses, of which 169 passed the qualifying questions. We removed three respondents whose answers contradicted each other, leaving 166. Consultants or advisors make up 8% of the 166, and investors under 1%; both may be answering about client or portfolio organizations.
One respondent's answer for Google's platform was recorded under Amazon Bedrock Agents and is counted as Google. Eight respondents named a primary platform they had not ticked as one they use. We leave them out of the share of primary picks, which gives a base of 154. In Finding 2's "Users who name it primary" column, they still count as users of the platforms they ticked, and as users who did not name those platforms primary.
For Anthropic, we named Claude Platform and Agent Skills in the platforms-in-use question, and Claude Agent SDK and Managed Agents in the platforms-considered question. The OpenAI and Google options named the same products in both questions. Some answer options carried descriptive labels in the questionnaire; we give each option's plain meaning, and keep the "model gravity" label where we quote that option.
Comparisons with the July Pulse use its 107 qualified respondents. Two July figures differ slightly from those published in July: the share running two or more platforms is 87% here, on the 105 respondents who named a listed platform, against 85% on all 107; and custom in-house orchestration is 21% when computed from its count, against the published 22%.
Some groups in this report have fewer than 40 respondents: Microsoft users, 31; Salesforce users, 25; Bedrock users, 15; planners considering Anthropic, 35; four Finding 3 subgroups, 26 to 39; and July's single-platform users, 14. Percentages for groups this small are less precise; for a group of 14, a result could differ by about 26 percentage points either way from the true figure.
We treat a month-to-month change or a difference between two groups as enough to call only when a statistical test gives p below 0.05; otherwise we call it too close to call. We use the two-proportion z-test, or Fisher's exact test when either group has fewer than 40 respondents. To rank two answers or call one a majority, we use an exact binomial test, or the exact McNemar test on multiple-answer questions. July's and August's respondents differ in industry, company size and purchasing role, so a change between months may reflect who answered.