Skip to the cards
Filed
Source
Christopher Smith
Loans
674

MCP and A2A Need Identity, Context and Operations to Scale AI Agents

MCP connects agents to tools; A2A coordinates work between them. Enterprise scale also depends on identity controls, shared context, process redesign and monitoring.

MCP and A2A Need Identity, Context and Operations to Scale AI Agents

Connecting AI agents is getting easier. Running them safely and reliably across an enterprise is a much larger task. The Model Context Protocol (MCP) gives an agent a standard way to access tools and data; Agent2Agent (A2A) supports communication and task handoffs between independent agents. Neither protocol, by itself, supplies the identity controls, trusted business context or operating practices needed to manage a large agent network.

The distinction matters as organizations move beyond experiments. Gartner predicts the average global Fortune 500 enterprise will run more than 150,000 AI agents over the next two years. Collibra’s 2026 Hallucination Tax Report says 76% of respondents have encountered critical roadblocks while trying to take AI initiatives from pilot to full production during the past 12 months. More connections can speed up work, but without controls and shared knowledge they can also spread incorrect results faster.

MCP connects agents to resources; A2A connects agents to each other

MCP addresses an agent’s connection to tools, data sources and services. That can include local files, enterprise databases and SaaS applications. Instead of building and maintaining a custom integration for every tool, teams can use a common way for an agent to access those resources.

A2A addresses a different problem: how independently built agents find one another, exchange tasks and pass back results. Scott Gerlach, cofounder and chief security officer at StackHawk, compares the distinction to a plumbing job. A plumber uses tools to move a sink; when the work requires drywall, painting or reassembling a vanity, the plumber brings in other specialists. MCP is the tool access, while A2A enables the handoff.

John Pettit, CTO at Promevo, describes MCP as a “secure USB port” for connecting an AI model to enterprise resources. He characterizes A2A as the negotiation protocol that lets a general contractor coordinate specialized subcontractors. Arpita Soni, a senior member of IEEE, makes a similar distinction: MCP is the vertical connection between one agent and its tools or data, while A2A is the horizontal connection between agents built with different frameworks or by different vendors.

That split is already appearing in products. Claude, OpenAI, Microsoft Copilot and other platforms expose customization through skills, connectors and plug-ins. App-building services including Bolt.new, Lovable, Pave, Replit and v0 can connect to MCP servers or offer built-in connectors. MCP clients are also available in development tools such as Claude Code, Cursor, Devin Desktop, Gemini CLI, GitHub Copilot and Kiro. The protocols are becoming familiar building blocks, but adoption alone doesn’t settle who may use them or how their work should be managed.

Give each agent a governed identity

MCP and A2A describe ways to present credentials, but they don’t define an organization’s access policy. IT teams still need to decide which agents can reach which resources, what actions they may perform and which other agents they may contact.

Kevin Paige, CISO at C1.ai, notes that MCP binds a token to a single server audience and prevents that server from passing it elsewhere. A2A treats peer agents as opaque and can stop at an AUTH_REQUIRED response, leaving the client to obtain credentials outside the protocol. The protocols provide connection patterns; the organization must supply the identity and authorization rules around them.

That work includes setting access levels for confidential and unstructured data, with permissions tied to the agent, the situation and the action. Andrew Hartnett, CTO at Bitwarden, says agents acting in sensitive systems should be managed as privileged, non-human identities. He points to secrets management for API credentials, keys and tokens, least-privilege boundaries, and human review for high-risk actions.

Build shared context agents can trust

MCP can connect agents to a business context layer: the information that explains the organization’s terminology, entities, relationships and operating rules. That layer may include a knowledge repository, a semantic layer defining business terms and knowledge graphs describing how business entities relate. Its usefulness depends on data quality, governance and controls over what agents do with the information they retrieve.

Felix Van de Maele, CEO and founder of Collibra, describes governance as covering both the inputs an agent receives and its outputs and actions. If teams don’t agree on what business terms mean, he argues, they can’t reliably trust the agent’s answers. For administrators, that makes shared definitions and reliable data sources part of the agent architecture, rather than cleanup to defer until after deployment.

Organizations have encountered a related problem with business intelligence: direct connections to data sources through ODBC or JDBC can produce scattered connections and isolated views. Scaling agent integrations the same way risks creating similar fragmentation, with each agent relying on its own sources and interpretation. Data fabrics can provide a basis for shared context, but teams still need to improve knowledge quality and create feedback loops in which experts validate agent activity.

Vivek Vaidya, cofounder and CTO at Kana, warns that adding MCP endpoints and A2A messages won’t help a system learn if every agent starts from scratch. His preferred direction preserves shared context, decision history and human judgment so later agents can build on prior work. Barr Moses, cofounder and CEO at Monte Carlo, likewise cautions that implementing the protocols in silos can increase the speed of bad responses. James Oh, experimental product lead at Outtake, says agents need reliable ways to decide which tools and other agents to trust as their network grows.

Redesign processes around collaboration

Many agents are currently assigned narrow tasks, which can make access and performance easier to manage. A larger opportunity comes when teams reconsider whole processes and let agents contribute their separate expertise. Oskar Konstantyner, chief product officer at Templafy, describes a shift from assigning a known task to a known capability toward agents working together to determine what a task requires.

Processes such as order-to-cash, employee onboarding and compliance reporting were designed before current AI capabilities existed. Daniel Meyer, CTO of Camunda, says those workflows are legacy processes because they weren’t built with this generation of AI in mind. A2A can support delegation across specialized agents, while MCP connects each agent to the resources it needs. Marc Sanfaçon, CTO and cofounder at Coveo, stresses that the systems behind those connections still need to provide relevant, current information and enforce permissions.

Prepare operations to monitor agent work

As agents hand work to one another, tracing their actions becomes harder and more useful. Observability helps teams diagnose failed handoffs, investigate accuracy problems, find security gaps and support audits. David Mytton, CEO and founder at Arcjet, identifies agent-to-agent work as a security challenge because it can make activity harder to see.

Kamesh Tumsi, chief product officer at Smarsh, says environments with strict control requirements need consistent role-based access controls, audit logs and chain-of-custody records for interactions. Those operational records are a starting point for agenticops: the practices through which site reliability engineers monitor agents, resolve incidents and investigate root causes.

The practical takeaway is to treat MCP and A2A as connection protocols within a larger operating design. Before expanding deployments, teams need defined agent identities and permissions, a governed context layer, processes suited to collaboration and monitoring that follows work across agents. Those capabilities will determine whether a growing agent network becomes manageable infrastructure or a faster way to multiply errors.

Source: Christopher Smith · www.infoworld.com

Filed alongside