Enterprise investment in AI agent orchestration is surging, with organizations rapidly consolidating onto major model-provider platforms and building out sophisticated control planes. Yet a stark gap has emerged between ambition and reality: 71% of enterprises admit that a quarter or fewer of their deployed “agents” are true multi-step orchestrated workflows, with the vast majority remaining simple single-prompt chatbot wrappers. This is the central finding of the latest VentureBeat Pulse Research survey, which examined agent orchestration across 101 enterprises with 100 or more employees, revealing a landscape where the infrastructure for advanced agents is being built well ahead of the agents themselves.
The survey, conducted in a single June 2026 wave, provides a directional snapshot of an industry in transition. While enterprises are placing large, strategic bets on orchestration platforms, tooling, and control architectures, the actual deployed portfolio tells a more modest story. The orchestration layer is real and maturing quickly, but most of the “agents” running on it are not yet doing the work that layer was designed to manage.
Anthropic’s Claude Leads a Platform Consolidation Driven by Model Gravity
When asked which agent orchestration platform they primarily use today, enterprises overwhelmingly named the major model providers. Anthropic’s Claude Platform and Agent Skills leads decisively at 40%, more than double any rival. Microsoft AI Foundry and Copilot Studio follows at 18%, with OpenAI’s Agents SDK and Responses API at 13%. Google’s Enterprise Agent Platform (8%) and Amazon Bedrock Agents (2%) round out the top tier, meaning the five major model platforms together account for roughly 80% of primary deployments. In contrast, open frameworks like LangChain and LangGraph (6%) and custom in-house builds (5%) remain marginal in this cohort.
The selection logic behind this concentration is clear. When asked what most influenced their platform choice, 21% of respondents cited “model gravity”—native alignment with a state-of-the-art base model—as the single largest factor. This explains why Anthropic leads: enterprises are choosing the orchestration environment that comes with the frontier model they have standardized on. But the next-tier factors complicate the picture. Flexibility across models and tools (17%) and ease of development (17%) indicate that enterprises also want to avoid being trapped by that choice, a concern that becomes critical later. Security and permissions (14%) and total cost of ownership (11%) round out a pragmatic buying logic, while raw performance (latency and memory) sits last at 4%.
Despite the strong preference for model-provider platforms, satisfaction is provisional. Respondents rate their platforms at an average of 3.94 out of 5, with “value for money” also at 3.94 and “ease of implementation” the weakest score at 3.85. Crucially, 96% of respondents plan to change their orchestration approach within the year. This is a layer enterprises tolerate more than they love—they are building on it today, but the search for something better is already underway.
Reliable Multi-Step Execution Is the Primary Success Metric
Enterprises judge orchestration success by a single, clear standard: does it reliably complete the work? Task completion reliability is the leading success metric at 32%, followed closely by multi-step workflow management at 28%. Together, these two factors account for 59% of responses, dwarfing developer productivity (17%), end-user experience (9%), and operational stability (9%). This reliability-first lens is precisely what makes the “chatbot trap” finding so pointed: enterprises define success as dependable multi-step execution, yet most of their deployed agents are not yet doing multi-step work at all.
The ambition, however, is to change that. When asked about major strategic moves in the next 12 months, three responses clustered near the top: 25% plan to increase investment in custom, in-house orchestration control planes; 24% will standardize on a single centralized framework; and 23% will expand agents from sandbox into production. This is a clear story of consolidation and operationalization. Enterprises want fewer frameworks, more production exposure, and more ownership of the control layer—an appetite that is shaping the architecture they are putting in place.
The Hybrid Control Plane Is the Dominant Architectural Expectation
By the end of 2026, a clear majority of enterprises (51%) expect their primary agent control plane to be hybrid—combining provider-native orchestration with an external layer they control. Only 6% expect to hand control to a provider-managed service outright. The reason for this deliberate hybridity is singular: vendor lock-in. When asked what they fear most if control lives inside a model-provider platform, 35% named vendor lock-in, ahead of security and permissioning limitations (28%) and inflexibility across models and tools (21%).
This finding echoes a broader enterprise posture: build on a provider’s platform, but decline to be governed entirely by it. The hybrid control plane is the architectural hedge against the lock-in enterprises most fear. The data also shows movement on this question. In an earlier April–May wave (n=145), only 34% expected a hybrid control plane, and 12% expected full provider management. By June, the conversation had shifted unmistakably toward keeping control. Lock-in has also risen as a top concern, trading places with security and permissioning limitations from the earlier wave. The worry about provider platforms appears to be maturing from whether they can be secured to whether they can be replaced.
The Chatbot Trap: 71% of “Agents” Are Still Single-Prompt Assistants
The defining finding of this research wave is the gap between orchestration ambition and deployed reality. When asked to assess their own portfolios honestly, 62% of enterprises say only 1–25% of their deployed “agents” are true multi-step orchestrated workflows, with 9% admitting that none are—every deployment is a chatbot or prompt wrapper. Combining these bottom bands, a full 71% of enterprises say a quarter or fewer of their agents are genuinely orchestrated. Only 10% have crossed the halfway mark, and just 3% report that 76–100% of their deployments are advanced, largely autonomous systems.
This is not a contradiction; it is a roadmap. The platforms, budgets, and control architectures are being put in place precisely because the orchestrated portfolio is still so thin. The trap is not evenly distributed. Splitting the sample by organization size, 77% of smaller enterprises (under 2,500 employees) say a quarter or fewer of their agents do true multi-step work, compared to 62% of larger enterprises. Larger organizations are meaningfully further into genuine multi-step deployment; the chatbot trap is, directionally, a mid-market condition.
Investment Flows to Workflow Tooling, Security, and Scaling—But Fiscal Control Lags
Budget priorities align with strategic intent. Agent workflow tooling leads planned investment growth at 34%, followed by security and permissions enforcement (25%) and infrastructure for scaling agents (20%). The money is going to the machinery that strings steps together dependably and hardens agents for production. In contrast, agent monitoring and debugging draws a smaller 11%.
Fiscal control, however, remains a significant weak point. When asked how they enforce cost control over agent token consumption, 32% of enterprises rely entirely on native platform controls—built-in budget caps and throttling. More concerning, 27% admit they have no real-time, programmatic way to stop a runaway agent before a budget-breaking bill arrives; they only learn of it from the logs afterward. The enterprises building custom gateways (23%) or using dynamic routing to arbitrage across low-cost models (19%) are the ones treating token burn as an engineering problem. Once again, a size-based split appears: 34% of enterprises under 2,500 employees exercise only reactive control of agent spend, compared to 20% of larger enterprises. The mid-market is running the least mature agents on the least instrumented budgets.
What This Means for the Enterprise AI Roadmap
The picture that emerges is one of deliberate, strategic infrastructure build-out running ahead of a still-nascent deployed portfolio. Enterprises have decided how they want to orchestrate agents—on model-provider platforms chosen for model gravity, judged by reliable multi-step execution, backed by a deliberately hybrid control plane designed to avoid lock-in. The investment is flowing to workflow tooling, permissions, and scaling. The strategy is to consolidate frameworks and push agents into production. The orchestration layer is being built.
But 71% of the agents running on it today are still chatbot wrappers. The open question for subsequent waves is whether the deployed reality closes the gap on the ambition—or whether the chatbot trap proves stickier than the roadmap assumes. For enterprise leaders, the implication is clear: build the orchestration layer with the expectation that it will eventually run genuinely autonomous, multi-step workflows. But plan for a transition that may take longer than current enthusiasm suggests. The infrastructure for the agent-driven enterprise is being constructed now, but the agents themselves are only just beginning to arrive.
