Beyond A2A and MCP: AGNTCY’s Vision for Composable AI Agents

AI agents are no longer confined to toy projects and research papers. They’re becoming foundational to enterprise workflows, customer experiences, and autonomous decision systems. But as adoption accelerates, a critical architectural problem has emerged: today’s agents operate in silos. Built on different frameworks with incompatible schemas, they struggle to communicate, share context, or collaborate reliably across boundaries.

That’s the challenge AGNTCY is designed to solve.

Originally open-sourced in early 2025 through Cisco’s Outshift, with LangChain and Galileo as core collaborators, AGNTCY quickly gained traction across the ecosystem. In July 2025, stewardship transitioned to the Linux Foundation, where neutral governance and a growing contributor base—now over 75 organizations—ensure both stability and scale. Formative members include Dell Technologies, Google Cloud, Oracle, and Red Hat, alongside contributors such as LlamaIndex, CrewAI, Komodor, Redis, Dynamiq, and Orium, all invested in advancing shared standards for agent-to-agent interoperability.

At the heart of AGNTCY are two core components: the Open Agent Schema Framework (OASF) and the Agent Connect Protocol (ACP). Together, they define how agents describe themselves, discover others, authenticate securely, and exchange information in real time. The vision is ambitious: a federated “Internet of Agents” that enables modular collaboration, much like the early web enabled hyperlinking across networks.

AGNTCY’s role complements other emerging standards: the Model Context Protocol (MCP) focuses on tool access, and A2A addresses secure task delegation, while AGNTCY provides the broader integration layer that unifies identity, discovery, composition, and monitoring under a shared open specification.

If you’re building autonomous workflows or integrating agents into customer-facing systems, understanding how these standards relate—and why AGNTCY matters—may shape how you design, secure, and scale your AI infrastructure.

The current generation of agent-based systems is powerful, but fragmented. Most agents today are built for a narrow task, using a specific runtime or orchestration library (e.g., LangChain, CrewAI, and Autogen, among others). While these agents can operate well in isolation, their composability—the ability to work alongside other agents or tools without tight coupling—is limited.

This becomes a problem at scale. Whether an organization is orchestrating through a master-agent model or coordinating a more distributed set of peers, the challenges are the same: re-implementing glue logic, building custom adapters, and sacrificing reliability for flexibility. There’s no shared schema for how agents describe their capabilities, no standardized way to locate and verify other agents, and no mechanism to establish context across agents built by different teams or vendors.

A composable agent ecosystem requires the same kind of open foundations that powered the early internet: shared identifiers, message formats, and trust mechanisms. Just as importantly, it must remain flexible enough to adapt as new standards emerge and orchestration models evolve. AGNTCY aims to deliver exactly that.

And the payoff is clear. Research from the MACH Alliance shows that enterprises already well along in their composability journey are twice as likely to succeed with AI deployments—77% achieving success compared to just 36% for those new to MACH. The lesson is simple: openness and modularity don’t just reduce risk, they accelerate adoption.

AGNTCY’s impact grows with the number and diversity of agents participating in the ecosystem. Some emerging use cases include:

• Cross-vendor agent orchestration: Teams can publish interoperable agents that can be discovered and used by others, without shared infrastructure or tight coupling.
• Scalable and performant orchestration: AGNTCY ensures coordination happens with low latency and reliability across diverse environments, so multi-agent workflows don’t collapse under real-world scale.
• Secure agent marketplaces: Think of this as an “npm for agents,” but with reputational scoring, credential validation, and usage analytics.
• Domain-agnostic agent networks: Beyond single verticals like commerce or finance, agents can connect across functions and industries. LangChain workflows (like LangGraph) hint at what’s possible, with agents dynamically discovering and coordinating mid-execution. Enterprises benefit from modularity, security, and the ability to reuse agent components without reinventing governance and integration every time.

For teams building AI infrastructure, AGNTCY opens new possibilities, but it also demands strategic clarity. And as a result, there are a few important considerations.

First, it’s important to adopt a layered approach. Use AGNTCY for agent discovery and connection; use MCP for tool access; use A2A for secure peer delegation. Second, design for interoperability. Agents should be defined with OASF-compliant metadata and designed to accept structured context. And finally—and perhaps most importantly—plan for governance. As agents proliferate, managing their trust, auditability, and lifecycle becomes critical. AGNTCY bakes these concerns into the protocol.

The initiative is still early, but in my opinion the momentum behind it is strong. Participating early gives teams a voice in shaping the spec— and a head start in building agent networks that won’t need to be rewritten when standards mature.

The rise of AI agents mirrors the early days of the internet: promising, chaotic, and fragmented. AGNTCY proposes a new foundation for inter-agent communication, and also for a shared fabric of discovery, connection, and collaboration.

Its greatest potential lies not in replacing MCP or A2A, but in binding them together. By defining open schemas, real-time protocols, and trust layers, AGNTCY brings the agent ecosystem one step closer to a true federated model where modularity, interoperability, and composability are the defaults, not the exceptions.

If you’re building AI systems today, or planning to in the near future, my advice is this: don’t just build smart agents, build ones that can connect, cooperate, and evolve.