Can A2A and MCP (and Other Open Protocols) Solve the Challenges Within Multi-Agent, Multi-Vendor Systems?

Artificial Intelligence (AI) agents—Large Language Model (LLM)-based systems that autonomously make decisions, set goals, and take actions—represent the next evolution beyond Generative Artificial Intelligence (Gen AI). They offer clear benefits across industries by automating tasks and processes, typically in human-in-the-loop scenarios where prompts guide limited, contextual responses. Agentic systems take this further. These architectures include multiple agents supported by tools for planning, memory, and coordination. They can handle multi-step workflows autonomously. For example, while an agent might answer Human Resources (HR) questions, an agentic system could complete an entire hiring process from job spec to budget planning, each step managed by a different agent. However, building agentic systems is not simple and they present major commercial and technical challenges:

• Multi-Vendor Fragmentation: Enterprises are likely to develop systems based on agents from multiple vendors depending on the application, task, business unit, and LLM requirements. This will create silos, inhibit agentic workflow management, and limit inter-agent communication.
• Security Risks: Expanded attack surfaces, potential agent compromise, and cross-boundary data access introduce serious cybersecurity and Intellectual Property (IP) concerns. In addition, as systems expand with heterogenous components, expect asymmetric security protocols that may or may not align with existing enterprise requirements. As systems grow, enforcing homogenous security standards becomes even more challenging—deploying security requirements across these domains will be inherently challenging.
• Orchestration Complexity: Coordinating agentic workflows across business units and systems is difficult, especially without clear orchestration standards and central visibility over which agents are performing which tasks. With multi-vendor, multi-domain agentic systems, this becomes even more challenging as each agent will be operated through a different platform—bringing us back to the single pane of glass challenge that distributed architectures have always struggled with.
• Lack of Traceability: Poor visibility in agent decisions impairs troubleshooting, auditing, and trust in outcomes. Building workflows around closed agents with opaque decision-making processes will be impossible for most enterprises. Agentic system actions, decisions, and output need to be traced, recorded, and audited. This includes interactions between agents, the decisions each agent makes, etc.
• Blame and Accountability: Determining responsibility when agentic systems go wrong is a critical, unresolved issue. This is especially challenging within heterogenous environments with multi-vendor solutions. Then add increased automation as humans move out of the loop, and this challenge will grow. As agentic systems become more autonomous this challenge will again grow. Insuring mission-critical use cases that utilize agentic systems will be very challenging without clear accountability.
• Reliability Concerns: Integrating agents (based on LLMs) will compound the hallucination rates across a multi-step process, creating significant risks, especially in mission-critical deployments. As multi-agent systems develop, more processes and tasks will be automated with less visibility and human controls.

Multi-vendor distributed agentic systems will likely be the de facto option for large enterprises. Enterprise Information Technology (IT) systems are already heterogenous environments with multi-cloud deployments—agentic systems will be the same with different vendors providing more effective solutions for various applications. To solve this challenge, technology leaders have developed open, complementary protocols to help align and integrate agentic systems:

• Model Context Protocol (MCP): Launched by Anthropic in November 2024, it acts as a universal interface that enables AI agents to connect to digital tools, content repositories, business applications, and development environments. By replacing fragmented, custom-built integrations with a single protocol, MCP allows for vertical integration between AI systems and enterprise data sources. It facilitates secure, two-way connections through MCP servers. Since OpenAI announced support for MCP in March 2025, integrating it across its product stack, adoption has accelerated. Microsoft has announced support, Amazon Web Services (AWS) has joined the steering committee, and Google DeepMind has also confirmed future support in upcoming Gemini models; while other notable players like IBM and Cloudflare are also supporting it. While MCP increases reliability and reduces integration friction, it also introduces new security concerns, as it relies on external servers to link models and data sources, thereby expanding the enterprise attack surface.

• Agent-to-Agent (A2A): Launched by Google in April 2025, it supports horizontal integration by enabling agents to communicate, collaborate, and delegate tasks across heterogeneous systems and vendors. It standardizes how agents exchange information and coordinate actions, simplifying integration and accelerating deployment of distributed agentic systems. Initially backed by over 50 technology partners (with strong participation from system integrators), A2A continues to gain industry momentum as Microsoft recently announced that it will be adopting it, alongside financial institutions and fintech companies.
• Agent Communication Protocol (ACP): Launched by IBM in March 2025 as an open-source project under the Linux Foundation, ACP enables agents to communicate within tightly integrated, local environments without the need for external services. Designed for secure, controlled deployments, ACP addresses the challenge of inconsistent agent interfaces by acting as a universal connector, making it easier to build multi-agent systems internally.
• Agent Gateway Protocol (AGP): Launched by the AGNTCY initiative in April 2025, AGP supports diverse communication patterns such as request-response, publish-subscribe, fire-and-forget, and streaming. Built on the gRPC framework, AGP enables secure and scalable communication between agents, making it well-suited for complex agentic systems requiring high-performance messaging capabilities.

Collectively, these protocols demonstrate growing momentum toward open, interoperable frameworks that enable multi-vendor, Agentic AI systems. However, the market remains in an early stage without consensus and adoption across the board—leaving space for competition, coopetition, or cooperation. There are several potential paths for how this protocol landscape might unfold:

1. Competition & Ecosystem Fragmentation: While open protocols are designed to align the ecosystem, in practice, they can lead to fragmentation. A prime example is the Agent2Agent (A2A) protocol, which aims to facilitate effective communication between AI agents. However, A2A is currently centered around Google and its partner ecosystem. While Microsoft’s decision to adopt the protocol is a notable step toward broader adoption, it doesn't prevent other players from developing or aligning around alternative protocols tied to their own ecosystems. This limits interoperability and dilutes the value of "open" standards. In this scenario, the size and cohesiveness of a protocol’s ecosystem may become a significant competitive advantage.
2. Formal Standardization: Another potential outcome is that these protocols follow a path similar to that of the telecoms industry, where standards are developed and maintained through industry-wide coordination. In this scenario, vendors would work together to enable seamless interoperability across platforms and tools. As each protocol is open-source, it will facilitate standardization, but the challenge remains speed. Developing and ratifying standards is slow and complex, requiring broad consensus and a lengthy adoption cycle. Given the rapid pace of innovation in AI, this route is unlikely to keep pace with market needs.
3. De Facto Protocol Dominance (Most Likely Scenario): The most probable outcome is that certain open-source protocols become de facto standards through widespread adoption by market leaders. As hyperscalers, model developers, and AI platform providers begin to accept and integrate these protocols, their usage becomes a practical necessity for any vendor looking to ensure interoperability with the dominant platforms. This mirrors how open-source models became the default foundation in Gen AI—driven by utility, network effects, and developer adoption, rather than formal standard-setting.

Common protocols will drive the shift from isolated agents to integrated, multi-agent systems—opening up significant monetization opportunities for vendors across the AI value chain:

1. Agent-as-a-Service (Agent-aaS): Today’s agent systems are siloed, with platforms building and operating agents for specific tasks through proprietary integrations. Protocols will enable modular, on-demand agents that enterprises can select and use like Software-as-a-Service (SaaS) apps. Vendors can “stitch” agents into enterprise systems via standard protocols—unlocking new revenue streams without building from scratch. For example, Siemens could monetize its partner ecosystem by offering agentic services Over-the-Top (OTT), rather than custom-building agents for every use case.
2. Orchestration & Audit: Interoperability allows vendors to offer orchestration platforms that provide centralized visibility and control over complex, multi-agent environments—essential for mission-critical and regulated industries. Industrial vendors could integrate these tools within Product Lifecycle Management (PLM) systems.
3. Multi-Agent Support for Data Platforms: Platforms like Databricks can host and manage multi-vendor agents, enabling cross-agent collaboration within their ecosystem. These vendors can also shift to Agent-aaS models, offering integrated agent solutions within their existing AI stacks.