Survey finds 71% of enterprises say fewer than a quarter of their deployed agents are multi-step workflows
Anthropic’s Claude leads as primary orchestration platform for 40% of enterprises, but most 'agents' remain chatbot wrappers and cost controls lag.
1 source · cross-referenced
- Anthropic’s Claude is the primary agent orchestration platform for 40% of enterprises, more than double any rival, followed by Microsoft (18%) and OpenAI (13%).
- 71% of enterprises report that a quarter or fewer of their deployed 'agents' are true multi-step orchestrated workflows rather than single-prompt chatbot wrappers.
- By the end of 2026, 51% expect a hybrid control plane (provider-native plus external orchestration), while only 6% expect to hand control to a provider-managed service.
- 27% of enterprises lack real-time programmatic control to stop a runaway agent before the bill arrives.
Enterprises are consolidating agent orchestration onto major model-provider platforms, with Anthropic’s Claude leading as the primary platform for 40% of organizations—more than double the share of Microsoft (18%) or OpenAI (13%). The choice is driven by the perceived gravity of the underlying model, reflecting a preference for platforms aligned with state-of-the-art capabilities.
Despite this consolidation, the reality of deployed agents lags far behind the ambition. A majority of enterprises (71%) report that a quarter or fewer of their deployed 'agents' are true multi-step orchestrated workflows, with only 10% having crossed the halfway mark. This 'chatbot trap'—where single-prompt assistants are labeled as agents—highlights a disconnect between the orchestration layer being built and the actual capabilities of deployed systems.
The orchestration layer is evolving toward a hybrid model, with 51% of enterprises expecting a hybrid control plane by the end of 2026 that splits control between provider-native and external systems. Only 6% expect to fully cede control to a provider-managed service, driven by fears of vendor lock-in (35%), which now ranks as the top risk over security or flexibility concerns.
Fiscal control over agent token consumption remains underdeveloped. More than a quarter of enterprises (27%) lack real-time programmatic mechanisms to halt a runaway agent before costs escalate, relying instead on reactive monitoring or native platform caps. This gap is more pronounced in mid-market organizations, where 34% exercise only reactive spend control compared to 20% of larger enterprises.
Investment priorities reflect the push toward production-grade orchestration. Agent workflow tooling leads planned spending (34%), followed by security and permissions enforcement (25%) and scaling infrastructure (20%). The focus on tooling and permissions underscores the operational challenges of moving agents from sandbox environments to production, where reliability and multi-step execution are the primary success metrics.
The survey, conducted in June 2026 with 101 enterprises of 100+ employees, captures a transitional moment. While 68% of enterprises plan to adopt a new, additional, or replacement orchestration platform within 12 months—the highest switching intent observed—the largest group of these movers has not yet shortlisted a candidate. This suggests that the current concentration on major platforms is provisional, and the market for agent orchestration remains unsettled despite rapid consolidation.
The findings also reveal a strategic tension: enterprises are standardizing on model-provider platforms for their perceived model gravity, yet simultaneously building hybrid control planes to mitigate lock-in risks. This dual approach—leveraging provider platforms while retaining external control—reflects an industry still defining the boundaries of agentic systems in production.
- Jul 15, 2026 · Hugging Face
Hugging Face details Shippy agent architecture and hosting platform
Trust79 - Jul 14, 2026 · arXiv cs.AI
Message format effects in multi-hop agent relays depend on relay capability tier, study finds
Trust79 - Jul 13, 2026 · arXiv cs.AI
GATS framework achieves 100% success rate on agent planning tasks with zero LLM calls during inference
Trust79