Two protocols, MCP and A2A, are gaining ground for connecting and collaborating AI agents. At scale, they are not enough: security, governance, context and supervision become essential. Gartner anticipates more than 150,000 agents per Fortune 500 company within two years, while 76% of organizations say they are struggling to move into production. Here’s what these deployments actually entail.
Guardrails: avoid speed errors and safety blind spots
Barr Moses warns that when siled, MCP and A2A can especially accelerate erroneous responses. Deploying agents imposes observability standards, with tracking of steps and actions to diagnose cooperations, improve accuracy, address vulnerabilities, and support audits. David Mytton believes the next security challenge will emerge when officers work among themselves, making visibility more difficult. In regulated sectors, Kamesh Tumsi points out, every interaction must meet strict requirements. These issues come up against an operational reality: 76% of companies report critical blockages in the transition from pilot to production, while Gartner projects more than 150,000 agents per Fortune 500 company within two years.
Identities and authorizations: what MCP and A2A do not decide
Protocols define how an agent presents identifiers, without establishing governance. It is up to organizations to define the identity of each agent, their access rights and their authorized contacts. Kevin Paige says that both MCP and A2A fall under the Linux Foundation’s Agentic AI Foundation, but neither specifies who can initiate a call: MCP binds each token to a single server public and prevents transits, while A2A stops at AUTH_REQUIRED and leaves obtaining proof outside the protocol. Policies must govern access to unstructured and confidential data, detailing what actions are permitted, in what circumstances and with what granularity. Andrew Hartnett points out that these flows become non-human identities managed by programmatic secrets and recommends treating any agent operating on sensitive systems as a privileged identity, with fair privilege and with human safeguards for risky actions.
Context layers: data quality, semantics and trust
The MCP is used in particular to connect agents to the enterprise context layer, which brings together shared business semantics and entity and relationship graphs. At scale, this layer must rely on high-quality data, non-negotiable governance, and real-time AI governance. Felix Van de Maele distinguishes between the governance of inputs (context, semantics) and that of outputs (what the agent does and what he is authorized to do), and recalls that in the absence of agreement on the terminology, confidence in the answers is impossible. BI deployments connected directly via ODBC/JDBC have created connection webs and siled views, even among small organizations; following this path exposes you to “AI debt”. Data fabrics provide a foundation, but the work is to raise the quality of knowledge and connect agent activity to expert validations. Vivek Vaidya points out that adding endpoints or messages does not create a learning system if each agent starts from scratch; he advocates preserving shared context, history of decisions and human judgment. Trust between business experts and agent results must be built, with the context layer remaining in continuous evolution. Barr Moses warns against the illusion that a protocol would cover the trust layer, and James Oh insists on reliable ways for agents to evaluate which tools and peers to trust.
We test and decipher the new AI tools every evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Specialized agents promote process transformation
Many agents are designed by tasks, which facilitates access control and performance monitoring. As deployments progress, opportunities emerge to rethink processes through agent collaboration. Oskar Konstantyner describes a shift where each agent brings its expertise and co-determines the task, beyond the orchestration of predefined capabilities. Processes such as order processing, onboarding or compliance reporting can switch to sequences carried by cooperating agents, whereas, recalls Daniel Meyer, these processes date from before current AI. This overhaul requires A2A to delegate and exchange results between independent agents. Marc Sanfaçon specifies that A2A standardizes these exchanges and that, combined, MCP and A2A help navigate complex processes, provided that the systems provide relevant and recent information and enforce rights.
What MCP and A2A do, and where they stop
Technical managers clearly distinguish the roles: the MCP is a toolbox or secure USB port which connects an agent to files, databases or SaaS without tailor-made gateway code; A2A ensures negotiation and coordination between agents, such as a general practitioner and its subcontractors. For Arpita Soni, MCP standardizes the vertical access of a single agent to tools and data, while A2A provides horizontal connection: discovery, negotiation of capacities and passage of tasks between agents, with a common language to communicate and delegate. These protocols lay the foundation, but scale requires additional layers: identities and authorizations, enterprise context, reengineering, and agenticops. Their adoption is accelerating: creation tools (Bolt.new, Lovable, Pave, Replit, v0) plug into MCP servers, AI editors (Claude Code, Cursor, Devin Desktop, Gemini CLI, GitHub Copilot, Kiro) embed MCP clients, and large models expose customizations via MCP and A2A. They should influence the management of agents in companies, even if their simple presence does not regulate orchestration or governance.

