Organizations are moving toward operating multiple platforms to orchestrate AI agents rather than handing the task to a single provider, according to an ongoing analysis of VB Pulse data covering 107 organizations. Some 85% of organizations use two or more tools, while 64% use three tools, and only 15% rely on a single platform. This diversity is driven not only by a desire to avoid vendor lock-in, but also by concerns about vendors’ security and permission-management capabilities, as well as AI teams’ desire to impose their own controls.
Microsoft AI Foundry and Copilot Studio lead the tools currently in use, appearing in 70% of the environments surveyed. OpenAI Agents SDK follows at 68%, then Anthropic’s Claude Platform at 47%. Respondents also use, to varying degrees, Google’s Enterprise Agent Platform, LangChain and LangGraph, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex. Some 22% also operate custom orchestration tools internally alongside vendor services.
A Multiplatform Future
Some 53% of organizations expect their primary “control plane” to be hybrid by the end of 2026. By contrast, 14% expect to rely on a provider-managed service, 13% plan to build a custom internal control plane, and 11% are betting on external platforms that separate the orchestration layer from model providers.
This architecture does not yet appear stable; more than two-thirds of respondents plan to change their platforms within one year. Some 15% said the change would occur within three months or less, 24% within three to six months, and 28% within six to 12 months. Claude Agent SDK stands out among the tools under consideration, being explored by 43% of participating builders. About one-third are also evaluating Google’s Enterprise Agent Platform, 31% are focusing on custom internal orchestration, and 25% are examining OpenAI options.
Despite this ongoing shift, overall average satisfaction with the platforms in use reached 4.17 out of 5. However, satisfaction fell to 3.91 out of 5 when evaluating ease of implementation, and to 3.63 out of 5 when evaluating value for money. This gap suggests that the problem is not necessarily a lack of tools, but the difficulty of integrating and operating them in a way that justifies their cost.
What Are Organizations Actually Funding?
Flexibility tops the platform-selection factors at 29%, followed by security and permission considerations at 17%, and production reliability and control over agent execution at 15% each. Only about 10% of respondents considered “model affinity,” or native compatibility with a leading foundation model, an important factor. Ease of development was cited by 8%, total cost of ownership by 4%, and performance in terms of latency and memory by just 2%.
This focus is reflected in spending categories: agent observability and debugging account for 31% of spending, compared with 30% for security and permission enforcement, and 19% for workflow tools. This represents a shift from an earlier VentureBeat survey wave, in which spending on workflow tools was the largest category.
Organizations’ operational priorities center on task-completion reliability at 30%, multistep workflow management at 27%, developer productivity at 23%, and operational stability at 13%. Only 7% of organizations placed user experience among their top priorities, suggesting that the infrastructure-and-control phase still precedes improvements to interfaces and usability.
The Visibility and Cost-Control Problem
The leading platform-selection concerns are security and permission-management constraints at 37%, followed by vendor lock-in at 23%, limited visibility and observability at 22%, and inflexibility in model and tool selection at 16%. Most urgently, roughly one in five organizations cannot stop the spending of a runaway AI agent in real time.
To address this, 30% of organizations rely on native controls, such as budget limits or throttling, while 25% have built custom gateways that act as intermediaries to intercept runaway agents. Another 25% use dynamic routing to shift heavy tasks to less expensive models. By contrast, 21% rely on reactive monitoring after a problem occurs, such as reviewing logs later, without an immediate kill switch.
Organization size does not appear to be a decisive factor in financial-control maturity: 18% of organizations with more than 10,000 employees still rely solely on reactive control, compared with 23% of smaller organizations. In practice, this means that having greater resources does not automatically guarantee tools that measure spending and stop it before it escalates.
True Agents Have Not Yet Become Widespread
Participants’ responses show that many systems described as “agents” are still basic assistants or chatbots, rather than autonomous systems that perform multistep work. Some 47% said that between 26% and 50% of their systems represent genuine orchestration, while 35% said the share is no more than 1% to 25%. Some 14% said that 51% to 75% of their systems are complex, multi-agent workflows, while only 2% said that 76% to 100% of their systems are advanced and semi-autonomous. Another 3% said they are still deploying chatbots only.
These findings align with the June wave of VB Pulse, in which 71% of respondents said that a quarter or fewer of their deployed agents could independently complete multistep work, while only one-tenth of respondents said they had deployed agents at scale. Therefore, the current trend represents not so much the arrival of autonomous agents at a mature stage as the construction of the control layers needed to accommodate them in the future. While organizations are making progress in building the infrastructure, the gap remains between calling a system an “agent” and its actual ability to independently execute interconnected tasks.