Why 'Bring Your Own Agent' Is Key to MCP Infrastructure

Enterprise AI is fragmented. Anthropic, OpenAI, and Google continue to compete for market share, but there’s no clear winner yet. Developer teams are constantly testing new models, agents, and AI features to get an edge in software engineering, and balancing that never-ending innovation with stability is challenging for CIOs.

AI is unusually volatile compared to historical IT cycles, and enterprises are already becoming multi-model and multi-agent. These environments are also configuring more and more Model Context Protocol (MCP) servers to let agents connect with external data, tools, and APIs. Organizations are quickly finding that securing and governing MCP usage at scale requires a new layer of infrastructure to operationalize it in practice.

The problem is that these new MCP infrastructure layers, including server generation, configuration, authentication, discovery, tool curation, and threat detection, could easily become bespoke to a single agentic platform, limiting an organization’s ability to switch providers.

But given how rapidly platforms and models change, MCP infrastructure should work regardless of what agents you’re using or what models are underneath them.

Below, we’ll review the evolution of the AI market and consider why “bring your own agent” is so important for MCP infrastructure. The agentic infrastructure platforms that succeed will be the ones that abstract underlying agents and LLMs and are built to be malleable. This separation-of-concerns philosophy is key to applying consistent authentication and guardrails across any agent in vogue, whether it’s Cursor, Windsurf, Copilot, Claude, or whatever’s next.

AI models are a moving target

AI models are constantly shifting, with no signs of slowing down. The large language models (LLMs) themselves are changing day to day, and developers are constantly using different models in their day-to-day work. F5’s 2026 State of Application Strategy Report found organizations use seven AI models on average.

Given models are fine-tuned for different tasks, multi-model approaches are on the rise. An early 2026 global survey of 100 companies by Andreessen Horowitz found 81% of enterprises use three or more AI model families in testing or production. Models are continually differentiated for specific tasks, whether it’s high-level reasoning, general-purpose chat, knowledge management, customer support, or programming. Some organizations also chain multiple models together to compare, critique, and refine responses.

Academic research and development into AI models is constantly churning out new advances, quickening the pace of commercial model changes. Just look at Anthropic’s release cycle as evidence: Claude Sonnet 4.6 launched February 17, 2026, Sonnet 3.7 was retired two days later, and Sonnet 4 was retired by June 15. In 2025-26 alone, Anthropic retired about ten models.

Models that look foundational today may be outmoded in months, or less. Models can also go offline due to regulatory actions, as we witnessed with the Claude Mythos 5 and Fable 5 temporary suspension. DeepSeek’s rapid emergence was another reminder of how quickly the AI market changes. All this equates to uncertainty and unusual volatility for enterprise IT.

AI agents are constantly shifting

Not only are models undergoing constant development, but agents are constantly shifting too. Claude Code has quickly become one of the most widely used coding AI assistants, yet is remarkably only a year and a half old. Before that, everyone was excited about Windsurf, Cursor, and Copilot. Now, OpenAI Codex is gaining steam, as is Goose, an open-source alternative that decouples the agent from underlying model providers.

Developers tend to move quickly to whichever agent grants the most value, so allegiance to specific platforms is low. As proof, JetBrains 2026 research found Claude Code use at work jumped from about 3% in spring 2025 to 18% in January 2026, while Cursor’s growth slowed. What matters now is output: as agentic development becomes common, GitHub says commits on its platform nearly doubled year over year, surpassing 1.4 billion per month.

Fluctuation will likely continue as engineering chases velocity and multi-agent architectures become more common. Stack Overflow’s May 2026 pulse survey found 17% of developers use multiple specialized agents, and OpenAI reports that its heaviest Codex users generate over 60 hours of work per day across multiple agents working in parallel.

There are downsides to agent sprawl. For one, it could lead to shadow agents, which feels probable given more than half of employees surveyed by Okta report using AI tools without explicit approval.

With more agents in use, the chances for scope overload and permission drift also increase. Beyond the security implications, as more AI tooling is acquired, the chances of wasting time re-implementing agents due to incompatibilities with surrounding tools or digital infrastructure also rise.

AI infrastructure must be adaptive

IBM surveyed 1,000 senior executives in 2026 and found that 71% say switching their primary AI vendor or model would be difficult, and 91% don’t fully understand their dependencies across AI vendors, models, and infrastructure. There are no clear winners in enterprise AI, and guessing “what’s next” is a losing game. The answer, then, is to reduce the cost of being wrong by architecting for reversibility.

As I wrote on InfoWorld recently, open infrastructure will define the AI era. With all the fluctuation going on in the model and agent layer, you can’t assume you’ll be using the same model or agent day to day. The surrounding AI-native infrastructure must be adaptable to this reality: designed to be model- and agent-agnostic, so enterprises can swap models and agents at the drop of a hat.

With this in mind, the industry is progressing toward vendor-neutral AI infrastructure. This is where neutral, MCP-compatible infrastructure shines. A new category of MCP infrastructure is emerging to abstract the governance and connectivity surrounding agent-to-tool interactions, providing functionality such as:

• Governing and monitoring agent tool access
• Authenticating agent-to-MCP connections
• Optimizing how agents connect and make requests to MCP servers and APIs
• Static security testing and runtime threat detection
• Usage monitoring for chargeback use cases

Such infrastructure should use standard protocols like MCP to remain interoperable and reusable regardless of what agents you’re using or what models are underneath them. In this way, the governance of agentic environments is treated as a separate, portable infrastructure concern, rather than being intrinsic to every individual agent platform.

MCP gateways should be agent and model agnostic

Most enterprises that need to govern MCP use are also juggling multiple AI platforms and various underlying LLMs. When you spin up MCP servers, ideally the servers and configurations for their use should be abstracted from the AI platform, or else you risk significant rework when switching LLMs and agent providers.

As such, MCP gateways should be built agent and model agnostic in order to appeal to multi-model, multi-agent realities. Then, you can plug your MCP servers and tools into whatever agent or model happens to be trending now or next year, and preserve repeatable settings across an MCP portfolio. In doing so, you avoid reinventing the wheel or replacing infrastructure with every model or agent change.

In traditional enterprise IT, you already have a separation of concerns across many areas: removing identity from applications, API management from individual APIs, observability from individual services, and networking from workloads. MCP gateways are arguably the next evolution, acting as an abstraction layer between the tumultuous AI market and the need to retain durable foundations.

MCP gateways like Zuplo’s are built to be flexible by design, letting you bring your own agent and deploy unified authentication, routing, and observability across MCP server interactions. By decoupling MCP server generation and governance from individual AI platforms, you retain the ability to adapt to frontier technology with more agility.