MCP and A2A alone won’t scale AI agents
InfoWorld ·
Two protocols define how AI agents connect to internal resources and collaborate with other AI agents. The Model Context Protocol (MCP) is well known because it’s promoted as a general protocol for connecting AI agents to resources and other AI agents. But that’s not completely accurate. It’s really the Agent2Agent protocol (A2A) that’s enabling AI agent orchestration . The confusion stems from how quickly platforms have adopted these protocols and marketed them as features. Users building their own AI agents with Claude, OpenAI, Microsoft Copilot, and other models are already being exposed to customizing skills, connectors, and other plug-ins that use MCP and A2A. Vibe coding and AI app-building tools like Bolt.new, Lovable, Pave, Replit, and v0 can connect to MCP servers and include built-in connectors. AI coding editors like Claude Code, Cursor, Devin Desktop, Gemini CLI, GitHub Copilot, and Kiro have MCP clients. These protocols will shape how enterprises scale and manage AI agents. Gartner predicts that the average global Fortune 500 enterprise will run more than 150,000 AI agents over the next two years. But according to Collibra’s 2026 Hallucination Tax Report , 76% have hit critical roadblocks while attempting to move from pilot to full production during the past 12 months. Reviewing the two protocols and which services are your responsibility to layer on top of them is key to scaling enterprise-ready AI agents. Differences between MCP and A2A Scott Gerlach, cofounder and chief security officer at StackHawk , offers this analogy to help technologists better understand MCP and A2A’s functions. “If you think about doing some bathroom remodeling and you call a plumber to move a sink, he’d use tools in his MCP toolbox to complete his work. But when it’s time to fix the drywall, repaint, and put the vanity back together, the plumber agent would use A2A to call other contractor agents to complete that work,” Gerlach says. Another approach to explaining MCP and A2A comes from John Pettit, CTO at Promevo , who says that IT teams must recognize that MCP and A2A solve two entirely different integration challenges. “MCP is the secure USB port for an AI agent, or a universal power adapter for enterprise AI, and it allows an AI model to safely plug into local files, enterprise databases, or SaaS tools without engineering teams having to write and maintain custom API middleware for every connection. A2A is the negotiation protocol, like a general contractor managing a team of specialized subcontractor agents,” Pettit says. Arpita Soni, a senior member of IEEE , says that MCP and A2A solve two different layers of the agentic stack. “MCP is the vertical connection and standardizes how a single agent accesses external tools, data sources, and APIs, replacing the old pattern of writing a custom integration for every tool an agent needs. A2A is the horizontal connection, and it standardizes how independently built agents discover each other, negotiate capabilities, and hand off tasks, regardless of which framework or vendor built them. MCP gives each agent its own hands and eyes (tools and context or capabilities), while A2A gives separate agents a shared language to talk to one another and delegate work,” Soni says. MCP and A2A are protocols, but scaling AI agents requires layers of capabilities developed on top of them. Here are four at the top of the list: identity and authorization, connecting AI agents to the business’s context layer, business process re-engineering, and developing agenticops capabilities. Managing identity and authorizations MCP and A2A define how agents present credentials but don’t govern them. You’re responsible for defining the agent’s identities, what resources they can access, and which agents they can converse with. “Both protocols sit under the Linux Foundation’s Agentic AI Foundation , with one giving agents a dial tone to tools, the other a dial tone to each other, but neither saying who is authorized to place the call,” says Kevin Paige, CISO at C1.ai . “MCP binds every token to a single-server audience and bars servers from transiting any other. A2A treats peer agents as opaque and halts at AUTH_REQUIRED , leaving the client to obtain that credential outside the protocol.” Organizations will need to address unstructured data governance policies , especially around confidential data. Policies and implementation should determine which agents, under what circumstances, and for which actions access is granted to resources and at what granularity. “From an enterprise security perspective, both protocols transform automated workflows into non-human identities (NHIs) that rely on programmatic secrets management to govern their API credentials, keys, and tokens,” says Andrew Hartnett, CTO at Bitwarden . “Regardless of protocol, any agent authorized to act within sensitive systems must be managed as a privileged identity under strict least-privileged boundaries backed by human-in-the-loop guardrails for high-risk actions.” Connecting agents to the context layer One primary use case for MCP is connecting AI agents to the organization’s context layer . The layer often includes a knowledge repository (a semantic layer that defines business terms) and knowledge graphs that represent business entities and their relationships. To scale AI agents, organizations need a context layer built on high-quality data sources , non-negotiable data governance requirements , and runtime AI governance around agents . At a recent Coffee With Digital Trailblazers , a LinkedIn Live event that I host, I spoke with Felix Van de Maele, CEO and founder of Collibra , about the hidden costs of confident AI . “AI governance has two sides: the inputs around context and semantics, and governing the output around what the agent actually does and is allowed to do. If we don’t agree on the business terminology, there’s no way we can trust the agent with the answer,” Van de Maele says. We faced similar information access challenges in the early days of deploying business intelligence platforms and directly connecting them to data sources with ODBC or JDBC. Even small organizations ended up with a tangled web of connections and siloed data views. Organizations scaling AI agents risk mounting AI debt if they follow a similar path. Data fabrics are the foundation for building a shared context layer across AI agents, but the hard work is improving knowledge quality and developing feedback loops that connect agent activity with expert validation. “If every agent still starts from zero, more MCP endpoints or agent-to-agent messages simply move work faster without creating a system that learns,” says Vivek Vaidya, cofounder and CTO at Kana . “The durable architecture is the one that preserves shared context, decision history, and human judgment across those interactions, so each agent can build on what the system already knows instead of repeating the same reasoning from scratch.” Organizations looking to scale AI agents have to build trust between the company’s experts and the results of automated or recommended agent actions. The context layer needs ongoing development as more agents, resources, and agent interactions are enabled. “What I’d watch for is utilizing either protocol and assuming they cover the trust layer,” says Barr Moses, cofounder and CEO at Monte Carlo . “MCP and A2A implemented in silos have the potential to increase the velocity of bad responses.” James Oh, experimental product lead at Outtake , adds, “As we ask agents to do more on our behalf, they need reliable ways to decide which tools and other agents to trust, because that trust is what makes a larger network useful.” Business process re-engineering Many AI agents are task-based, which helps control information access and monitor performance. As organizations roll out more AI agents, they should consider strategic opportunities to re-engineer business processes through collaborations between agents. “Instead of one agent knowing exactly what to ask another to do, each agent brings its own expertise, and they work together to determine what the task actually requires,” says Oskar Konstantyner, chief product officer at Templafy . “It goes from orchestrating known capabilities to intelligent agents actually collaborating.” During another episode of Coffee With Digital Trailblazers, we discussed how AI agent collaboration will drive a wave of process re-engineering . Consider more complex business processes like order-to-cash, employee onboarding, and compliance reporting, where AI agent collaboration can replace multistep processes. Daniel Meyer, CTO of Camunda , said, “All of those business processes were designed in the world before AI existed. They’re all legacy because they were designed without the current generation of AI in mind.” So while the current focus may be on deploying AI agents and using MCP to connect them to the context layer and other resources, re-engineering business processes will require agents collaborating over A2A. “A2A standardizes how independent agents exchange tasks and results, letting agents delegate work so each part of a process goes to the agent best suited for it,” says Marc Sanfaçon, CTO and cofounder at Coveo . “Together, MCP and A2A help specialized agents work through complex business processes, but the systems behind them still need to deliver relevant, up-to-date information and enforce permissions.” Planning for agenticops A key consideration when rolling out AI agents is implementing them with observability standards . Tracking steps and actions will be needed to diagnose agent collaboration, improve accuracy, address security gaps, and support audits. David Mytton, CEO and founder at Arcjet , says, “The next big security challenge is when agents start doing work for each other, and it gets even more difficult to see what’s happening.” Kamesh Tumsi, chief product officer at Smarsh , adds, “In regulated environments, every interaction is subject to the same role-based access controls , audit logging, and chain-of-custody requirements, making it the only integration path that doesn’t create new compliance exposure.” Observability is the first step toward building an agenticops capability, where site reliability engineers will have the expanded role of monitoring, resolving issues, and performing root-cause analysis around AI agents. If enterprises have 150,000 AI agents running in the next two years, expect improvements in MCP and A2A protocols and a need to build out identity management, AI context governance, business process re-engineering capabilities, and agenticops.