New VB Pulse information exhibits Microsoft and OpenAI main enterprise agent orchestration, however Anthropic’s first measurable foothold factors to a bigger combat over who controls the infrastructure the place AI brokers run.
For the final two years, the enterprise AI race has principally been framed as a mannequin struggle: OpenAI’s GPT collection versus Anthropic’s Claude versus Google’s Gemini, with smaller and open-source alternate options additionally coming in from the U.S. and China.
However the subsequent strategic combat might not be over which mannequin solutions a immediate greatest. It could be over who controls the layer the place brokers plan, name instruments, entry information, run workflows and show to safety groups that they didn’t do something they weren’t presupposed to do.
New VB Pulse survey data suggests the class is already taking form. Our impartial Enterprise Agentic Orchestration tracker, a survey that data the preferences of certified, verified technical-decision maker respondents at enterprises at common intervals, discovered that Microsoft Copilot Studio and Azure AI Studio led with 38.6% primary-platform adoption in February, up from 35.7% in January.
OpenAI’s Assistants and Responses API held second place, rising from 23.2% to 25.7%.
Anthropic remained far smaller, but it surely made its first look within the tracker: shifting from 0% in January to 5.7% in February for Anthropic instrument use and workflows.
The underlying transfer is small — 4 respondents out of a complete 70 on this cohort, with extra to come back — however strategically fascinating as a result of it marks the primary signal on this tracker of Claude utilization shifting from the mannequin layer into native orchestration.
That distinction issues. Enterprises usually are not merely selecting chatbots. They’re deciding the place the dwell operational equipment of AI work will sit: inside Microsoft’s stack, inside OpenAI’s API layer, inside Anthropic’s managed runtime, inside an open framework, or throughout a hybrid mixture of all of them.
“That is the convergence second for enterprise AI,” mentioned Tom Findling, CEO and cofounder of AI cybsersecurity startup Conifers, in an announcement to VentureBeat. “Fashions and agent frameworks have matured sufficient collectively that enterprises are actually shifting focus past mannequin high quality to the management aircraft round it. In safety operations, we’re seeing the aggressive benefit transfer towards platforms that may orchestrate brokers, leverage enterprise context, and supply governance and auditability throughout buyer environments.”
Anthropic’s quantity continues to be small to start out — however the improve shouldn’t be
The Anthropic quantity, by itself, shouldn’t be overread. A transfer from zero to five.7% shouldn’t be a juggernaut. It isn’t proof that Anthropic has captured enterprise orchestration.
It isn’t even sufficient to say Anthropic has a sturdy lead in any a part of this market. Microsoft owns the early enterprise distribution benefit, and OpenAI has a a lot bigger put in base in orchestration than Anthropic.
However small numbers can matter after they seem at the beginning of a brand new market construction. Anthropic’s emergence in orchestration comes because the broader VB Pulse information exhibits Claude additionally gaining huge enterprise adoption on the mannequin layer.
In our VB Pulse Q1 Basis Fashions and Intelligence Platforms tracker, Anthropic rose from 23.9% in January to twenty-eight.6% in February after which much more dramatically to 56.2% in March amongst certified enterprise respondents, with the March studying flagged as directional solely, as a result of the pattern was solely 16 respondents.
The story, then, shouldn’t be that Anthropic is successful orchestration at the moment. It’s that Anthropic’s mannequin momentum could also be beginning to spill into the orchestration layer.
That’s the place the strategic stakes get greater.
A mannequin is less complicated to swap than an agent runtime
A mannequin is comparatively simple to swap, at the very least in principle. An organization can route one workload to Claude, one other to GPT, one other to Gemini and one other to a smaller open mannequin.
The truth is, the VB Pulse Basis Fashions tracker over the identical Q1 interval exhibits that multi-model technique is the enterprise consensus: respondents more and more report adopting a number of fashions and constructing orchestration layers that route throughout them by job, value and danger profile.
An agent runtime is completely different. As soon as an organization’s workflows, instrument permissions, credentials, audit logs, reminiscence, sandboxed execution and operational monitoring dwell inside one supplier’s atmosphere, switching suppliers turns into much less like altering fashions and extra like altering infrastructure.
That’s the actual purpose Anthropic’s 5.7% foothold is value watching
Anthropic has already made clear that it desires to offer greater than the mannequin. Its Claude Managed Brokers documentation describes a public beta for a managed agent harness with safe sandboxing, built-in instruments and API-run classes, whereas Anthropic’s engineering publish frames the structure round decoupling the mannequin from the encompassing agent equipment: the session, the harness and the sandbox.
In plain English, Anthropic is making an attempt to host the atmosphere the place Claude brokers keep in mind context, use instruments, run code, function inside sandboxes and persist throughout long-running workflows. That’s now not simply inference. That’s operational infrastructure.
The pitch is apparent: most enterprises don’t need to sew collectively their very own agent stack from scratch. They need brokers that may act, however additionally they need permission boundaries, audit trails, workflow reliability and methods to cease the system when one thing goes incorrect.
Safety is turning into the shopping for criterion
The VB Pulse orchestration tracker exhibits that patrons are prioritizing precisely these considerations. Safety and permissions ranked as the highest orchestration platform choice criterion in each January and February, at 39.3% and 37.1%.
Management over agent execution rose from 17.9% to 22.9%, whereas flexibility throughout fashions and instruments fell from 35.7% to 25.7%. The market seems to be shifting from optionality towards governance.
That shift isn’t a surprise. A chatbot may be incorrect and nonetheless stay principally contained. An agent that may ship emails, modify paperwork, question databases, name APIs or execute workflows has a a lot bigger blast radius. The enterprise query shouldn’t be solely whether or not the agent is sensible sufficient.
It’s who gave it permission, what it touched, what it modified, whether or not these actions had been logged, and whether or not the corporate can unwind the harm if one thing goes incorrect.
Ev Kontsevoy, cofounder and CEO of Teleport, an identification and digital infrastructure options firm, argues that the trade continues to be placing an excessive amount of emphasis on orchestration itself and never sufficient on identification: “The race to personal the agent orchestration layer is actual,” Kontsevoy mentioned. “It’s additionally fixing the incorrect downside first. Orchestration with out identification solely multiplies chaos. With out identification, you don’t know what an agent can entry, what it really did, or how one can revoke its entry when it operates outdoors coverage. A unified identification layer is a prerequisite to deploying brokers — one or many — in infrastructure.”
Syam Nair, Chief Product Officer on the clever information infrastructure firm NetApp, believes information administration is essential in all instances to safe AI agent orchestration throughout the enterprise. As he mentioned in an announcement to VentureBeat: "Efficient agent administration requires built-in intelligence and a repeatedly up to date understanding of each information and, critically, its metadata. This visibility permits organizations to outline and implement clear insurance policies so information is used solely by the suitable brokers, for the suitable functions. Making this work at scale is a crossfunctional effort. Safety, storage, and information science groups should work collectively to implement insurance policies that safeguard firm information, whereas creating a robust information basis for AI."
He continued: "The CIOs and expertise leaders which might be profitable are those who take the enter, insurance policies, and imaginative and prescient from all these groups under consideration as they construct a knowledge infrastructure that minimizes danger and drives enterprise worth."
Microsoft has the distribution edge
That’s the reason Microsoft’s early lead is sensible. Copilot Studio and Azure AI Studio sit inside an enterprise stack many firms already use: Microsoft 365, Groups, Entra ID, Azure and present procurement relationships.
The VB Pulse Orchestration Tracker for Q1 2026 describes Microsoft because the enterprise default, with no different platform inside 13 proportion factors in February.
David Weston, CVP, AI Safety, Microsoft, supplied some perception on why, writing in an announcement to VentureBeat: "With out a unified management layer, you begin to see fragmentation – brokers working in silos, inconsistent governance, and gaps in safety. What clients are asking for is a strategy to carry order to that complexity. With Agent 365, we’re offering a single management aircraft to look at, govern, and safe brokers throughout Microsoft, associate, and third-party ecosystems, all grounded in enterprise information and identification."
OpenAI’s second-place place can be unsurprising. Its Assistants and Responses API gave builders an early strategy to construct agent-like methods utilizing OpenAI’s fashions and tooling. Within the orchestration tracker, OpenAI shouldn’t be surging, however it’s nonetheless ticking up steadily: 23.2% in January to 25.7% in February.
Anthropic is the newcomer on the orchestration layer. However its timing could also be favorable. The VB Pulse Basis Fashions tracker for Q1 2026 suggests enterprises more and more see Claude as a match for higher-stakes workloads the place security, instruction following, lengthy context and governance matter.
The orchestration tracker suggests those self same patrons are actually shifting from agent experiments towards manufacturing workflows, the place safety, permissions and job reliability change into the gating points.
That creates a potential path for Anthropic: to not beat Microsoft because the default enterprise platform, at the very least not instantly, however to change into the agent runtime for firms that already belief Claude for delicate or advanced workloads.
The chance is lock-in
The chance for enterprises is lock-in.
The orchestration tracker discovered that a hybrid management aircraft — combining provider-native orchestration with exterior orchestration — was the main anticipated structure, holding round 35% to 36% throughout the 2 substantive waves.
Supplier-managed-only approaches grew modestly however remained a minority. The report’s conclusion is blunt: enterprises usually are not prepared to provide full orchestration management to any single supplier.
It makes complete sense as enterprises search to leverage the "best-in-breed" fashions, harnesses, and instruments from a number of distributors, particularly as their wants differ extensively throughout sector, enterprise, and measurement.
"Most enterprises will function in a multi-model, multi-agent atmosphere, which makes an impartial management aircraft important," agreed Felix Van de Maele, CEO of Collibra, an organization specializing in unified governance for information and AI, in an announcement to VentureBeat. "That’s the reason we constructed AI Command Middle: to provide organizations the visibility, governance, and real-time oversight wanted to handle AI methods and brokers throughout the total lifecycle."
That warning exhibits up within the danger information. When requested about dangers if agent management lives inside a mannequin supplier platform, respondents cited safety and permissioning limitations as the highest concern. Vendor lock-in was the second-largest concern and the just one that elevated from January to February, rising from 23.2% to 25.7%.
That is the stress on the coronary heart of the agent market. Enterprises need managed infrastructure as a result of constructing dependable brokers is difficult. However the extra a supplier manages, the extra it could personal.
Dr. Rania Khalaf, chief AI officer at WSO2 — the subsidiary of EQT that provides open supply, customizable AI stacks for enterprises — mentioned enterprises will want an agent management aircraft that sits other than particular person frameworks, harnesses and runtimes as a result of brokers mix the unpredictability of LLMs with the flexibility to take actions which have penalties.
“Groups need the liberty to make use of the most effective mannequin and framework for every job — Claude for coding, Gemini for writing, LangGraph or CrewAI for dynamic modular habits — and that heterogeneity makes constant governance untenable in built-in platforms that lock into one ecosystem,” Khalaf mentioned.
From LLMOps to Agent Ops
Khalaf mentioned the trade can be shifting from MLOps to LLMOps to “Agent Ops,” the place governance has to cowl the entire agent, not simply the mannequin name.
“A guardrail on an LLM name can catch hallucination or poisonous output, but it surely won’t catch an agent thrashing in an unbreakable, pricey loop, which is why governance now has to increase out from the LLM interplay to the scope of the agent,” she mentioned.
The sensible implication is that enterprises must separate coverage and management from the agent logic itself. Khalaf pointed to the latest instance of an agent deleting a manufacturing database regardless of being instructed to not, arguing that the failure confirmed the bounds of counting on prompt-level directions the place exhausting identification and entry controls are wanted.
“Pulling guardrails, evals, insurance policies, bindings, and agent identification out of the core agent logic permits them to be configured per deployment and per atmosphere, owned by the suitable groups in safety, product, and compliance, with out fragmenting the governance layer as completely different groups select completely different fashions and frameworks,” Khalaf mentioned.
MCP is open. The runtime should be sticky
That’s the place Anthropic’s Mannequin Context Protocol, or MCP, complicates the story. MCP shouldn’t be a walled backyard; Anthropic launched it as an open normal for connecting AI methods to information and instruments, and Anthropic’s documentation describes MCP as an open-source normal for connecting AI purposes to exterior methods.
However openness on the protocol layer doesn’t routinely eradicate lock-in on the runtime layer. An enterprise may use an open protocol to attach instruments whereas nonetheless turning into depending on a supplier’s managed classes, logs, sandboxes, permissions mannequin, workflow state and deployment atmosphere. In different phrases, MCP might cut back integration friction, whereas managed agent infrastructure may nonetheless improve switching prices.
Khalaf mentioned Microsoft’s lead possible displays its M365 and Azure distribution, whereas Anthropic’s rising foothold may mirror a unique architectural wager round open protocols equivalent to MCP. However she argued the long-term path shouldn’t be a single-provider stack.
“Enterprises severe about operating brokers in manufacturing will find yourself multi-vendor throughout these layers,” Khalaf mentioned, “which is why the open and interoperable management aircraft issues greater than the present percentages may counsel.”
The following cycle could also be cross-vendor collaboration
That very same pressure — between provider-native comfort and cross-vendor actuality — is the place Arick Goomanovsky, CEO and cofounder of common AI agent orchestrator startup BAND, sees the following aggressive cycle forming.
“Enterprises now run brokers all over the place: particular person assistants and coding brokers, multi-agent methods in manufacturing, brokers embedded in Agentforce and ServiceNow, and third-party brokers consumed as agent-as-a-service,” Goomanovsky mentioned. “None of them collaborate throughout these boundaries by default.”
Goomanovsky argues that the lacking layer is not only orchestration inside a single mannequin supplier, however a cross-vendor collaboration layer that lets brokers from completely different ecosystems act collectively.
“What’s rising in parallel is demand for an agentic collaboration harness – an interplay layer that lets brokers from Microsoft, OpenAI, Anthropic, and inner groups function as one workforce,” he mentioned. “Orchestration inside any single vendor continues to be a walled backyard so the following aggressive cycle is cross-vendor agent collaboration.”
Impartial frameworks face an enterprise packaging downside
There may be additionally a warning signal for impartial orchestration frameworks. LangChain and LangGraph fell from 5.4% to 1.4% as the first orchestration platform within the certified enterprise pattern.
Exterior orchestration abstracted fully from mannequin suppliers additionally fell from 8.9% to 2.9%.
Scott Likens, World Chief AI Engineer at skilled providers large PwC, has a entrance row seat to this pattern as the corporate spearheads and assists purchasers with their AI transformations.
As he instructed VentureBeat in an announcement: "Proper now, most enterprises are nonetheless working in fragmented environments, with orchestration unfold throughout platforms, enterprise purposes, and internally developed tooling. Over time, the market will possible transfer towards extra unified orchestration fashions, however interoperability, governance and safety will stay important as a result of enterprises are unlikely to standardize on a single agent ecosystem."
The report argues that absolutely impartial orchestration frameworks might not but have the enterprise packaging — safety certifications, help, compliance documentation and vendor accountability — that procurement groups require.
That doesn’t imply open frameworks are irrelevant. It does counsel that enterprise patrons might more and more eat open or developer-first orchestration by managed merchandise, cloud-provider partnerships or inner management planes reasonably than as standalone frameworks.
The agent market begins to seem like cloud infrastructure
That is the place the agent market begins to look much less just like the early chatbot market and extra like enterprise cloud infrastructure. The successful distributors won’t solely have succesful fashions. They’ll have identification integration, permission controls, audit logs, observability, workflow tooling, sandboxing, analysis and a reputable reply to who owns the management aircraft.
Certainly, the orchestration layer is however one a part of the stack that the enterprise should fill in, and enterprises may very well determine to have completely different orchestration layers for brokers working in numerous departments and capabilities.
As Nithya Lakshmanan, Chief Product Officer at income group AI orchestration startup Outreach.ai wrote in an announcement to VentureBeat: "Normal-purpose orchestration platforms coordinate agent exercise effectively, however they don't carry the workflow-specific context that determines whether or not an agent's motion is right for a given state of affairs. In income workflows, an agent performing on incomplete deal historical past or lacking purchaser context will underperform and erode belief with customers. The groups getting essentially the most out of multi-agent methods are treating domain-specific information because the governance layer, with orchestration sitting on high. Most enterprises have chosen their orchestration stack, and what they're now determining is how these platforms get entry to the workflow context they should make brokers helpful inside particular enterprise capabilities."
That’s the reason Anthropic — which is more and more launching its personal domain-specific agents for finance and design, amongst different classes — is value following carefully. The corporate doesn’t must win all the orchestration market tomorrow for its technique to matter. It solely wants to influence a rising set of Claude enterprise clients to let Anthropic deal with extra of the encompassing equipment: instruments, workflows, reminiscence, execution and governance.
If it succeeds, Claude turns into greater than a mannequin in a multi-model portfolio. It turns into a part of the infrastructure the place enterprise work will get performed.
That will put Anthropic in a extra direct combat with OpenAI and Microsoft — not simply over mannequin high quality, however over the working layer of AI brokers.
The slender however vital learn
The protected interpretation of the VB Pulse information is slender however vital: Anthropic shouldn’t be but a serious enterprise orchestration platform. Microsoft is. OpenAI is far nearer. However Anthropic has registered its first measurable foothold on the orchestration layer, simply because the market is deciding who ought to management agent execution.
For enterprise patrons, which may be the query that issues most in 2026. Not which mannequin is greatest, however which supplier will get to run the agent — and the way exhausting it is going to be to go away as soon as the agent is operating.