Engineering teams are increasingly juggling multiple orchestration layers to mitigate these financial and security risks.
The Multi-Orchestration Reality in Enterprise Tech Stacks
Enterprise AI teams have systematically stopped betting on a single orchestration platform. The median enterprise now runs three distinct orchestration layers simultaneously. This architectural choice is driven by a deep-seated distrust of vendor-locked security boundaries and closed permissioning capabilities.
Data from the ongoing VB Pulse analysis of 107 enterprise tech stacks reveals a fragmented, highly plural ecosystem:
- Microsoft AI Foundry / Copilot Studio: Deployed in 70% of enterprise stacks.
- OpenAI Agents SDK: Deployed in 68% of enterprise stacks.
- Anthropic Claude Platform: Deployed in 47% of enterprise stacks.
Additional usage spans Google’s Enterprise Agent Platform, LangChain/LangGraph, Salesforce Agentforce, Amazon Bedrock, and LlamaIndex. Furthermore, 22% of engineering teams run custom in-house orchestration code to augment third-party tools. More than half of respondents—53%—expect their primary control plane to be fully hybrid by the end of 2026. Only 14% plan to rely on a single provider-managed service, 13% are building custom in-house control planes, and 11% are putting their capital into external abstraction platforms.
Platform churn is occurring at a rapid clip. Over two-thirds of surveyed builders plan to change their foundational platforms within a twelve-month window. Fifteen percent intend to migrate within three months, 24% within three to six months, and 28% within six to twelve months. Among platforms under active consideration, Anthropic’s Claude Agent SDK leads with 43% of builders exploring its capabilities. Roughly one-third are evaluating Google’s Enterprise Agent Platform, 31% are looking at custom orchestration, and 25% are investigating OpenAI’s ecosystem options.
Budgetary Allocation and the Pursuit of Control
Enterprise buying logic has pivoted decisively away from raw model performance toward operational governance. When evaluating orchestration platforms, flexibility leads procurement decisions at 29%. Security and permissions follow at 17%, while production reliability and execution control each claim 15%.
By contrast, model gravity—meaning native alignment with a state-of-the-art base model—matters to just 10% of buyers. Ease of development sits at 8%, total cost of ownership at 4%, and raw latency or memory performance accounts for a mere 2%.
Capital expenditure directly mirrors these priorities. Engineering leadership is investing the heaviest budget into agent monitoring and debugging tooling at 31%, closely followed by security and permissions enforcement at 30%. Workflow tooling commands another 19% of the budget. This represents a distinct structural shift from earlier deployment waves, where workflow tooling dominated spend outright.
Satisfaction metrics collected from builders highlight friction points in current deployments. While overall platform satisfaction scores a solid 4.17 out of 5, ease of implementation drops to 3.91 out of 5. Value for money sits lower still, at 3.63 out of 5.
The Visibility Deficit and Runaway Agent Budgets
Underneath the deployment metrics lies a severe operational vulnerability: runaway token consumption and token burn. One in five enterprises cannot stop an autonomous AI agent’s spending in real time. Without instantaneous programmatic interrupts, runaway loops can drain API budgets within minutes.
To curb unpredictable expenditures, organizations are implementing varied defensive strategies:
- Native Platform Controls: 30% of teams rely on built-in budget caps and API throttling provided directly by vendors.
- Custom Gateway Plumbing: 25% have engineered proxy middleware to intercept and inspect agent requests before they hit upstream LLM endpoints.
- Dynamic Routing: 25% use intelligent routing logic to offload heavy multi-step tasks to lower-cost, smaller-parameter models.
- Reactive Monitoring: 21% still depend entirely on post-hoc log analysis, meaning these enterprises possess zero real-time programmatic kill switches to halt live agent loops.
Organization size offers no immunity against this fiscal blind spot. Among enterprises with over 10,000 employees, 18% still exercise only reactive control over their token spend, compared to 23% of smaller organizations.
Beyond the Chatbot Wrapper
Despite heavy investment in infrastructure, true multi-step autonomous agency remains rare across the corporate landscape. Self-assessment data from builders shows that chat-centric interfaces still dominate daily operations. A small number of respondents report that 76 to 100% of their deployed systems are advanced and largely autonomous. Fourteen percent state that 51 to 75% of their systems consist of complex, multi-agent pipelines. Nearly half—47%—report that true orchestration handles between 26 and 50% of their workloads. On the lower end, 35% say orchestration accounts for just 1 to 25% of their systems, while 3% are running basic chatbots devoid of agentic capabilities.
Enterprises are constructing the multi-vendor control planes required for tomorrow’s autonomous workflows. Until real-time observability and instantaneous budget throttling become standard architectural defaults, keeping runaway agent spending under control remains a high-stakes engineering challenge.
