AI agents are proliferating as enterprises automate more of their operations. But ensuring agents can accomplish their tasks properly and securely remains a problem without an obvious solution.
Agents built to run in different environments typically don’t work well together. For instance, one organization might have a commercial off-the-shelf agent and another might have built a bespoke agent. They can’t communicate without some code to stitch them together.
“How do you get those things to work together?” asked Arnal Dayaratna, research VP of software development at IDC. “No one really knows, so part of this involves the invention of a shared data layer that allows agents to collaborate.”
For now, experts say, IT leaders must keep their AI infrastructure and architectures modular and flexible, in order to pivot when the market changes — and the market is always changing.
When communication breaks down
AI agents built today rely on special protocols, or communication methods, that imbue them with the structured vocabulary needed to break complex goals into manageable sub-tasks and hand them off. Naturally, a few protocols have emerged.
The most prominent is Model Context Protocol. Built by Anthropic, MCP determines how agents access resources, tools and information, such as APIs or databases. MCP is primarily optimized for agents within a shared trust boundary or a single application stack, such as Claude Desktop.
Google’s Agent2Agent focuses on how agents communicate with each other, coordinate tasks and collaborate on workflows. A2A agents advertise themselves using a structured metadata format called agent cards, a profile of sorts that enables agents to discover each other.
On the fringes is Agent Network Protocol, an open-source offering that supports open web agent discovery.
The kicker: Each protocol was built on fundamentally different assumptions about where agents live and who they trust. And they weren’t designed to talk to each other natively, according to Microsoft research. This means that organizations must build APIs to bridge the protocols.
Building a bridge as the market evolves
The bridge-building is happening. However, it may not keep pace with new innovations and evolutions in agentic technology. Few agents today share an access layer that helps maintain the identities and consistency large enterprises require to honor governance.
Moreover, the agentic protocols of today won’t necessarily be viable tomorrow.
“All of these will get overturned and subsumed and it’s going to happen super fast,” said Exabeam Chief AI Officer Steve Wilson. For example, Wilson said MCP is falling out of fashion due to its complex, server-based experience, while developers want a simple text file that enables them to build a skill.
Most IT leaders tend to err on the side of caution when it comes to high-stakes technical issues. Yet IT leaders must experiment with AI applications using disparate protocols, albeit leaving themselves the flexibility to course correct quickly.
“Given how fast things are evolving, if you’re not experimenting and pushing to see where things are at, you can get left behind quickly,” said Johnson & Johnson CIO Jim Swanson.
To that end, the speed of the evolving AI market led J&J to apply modular architectures and reimagining business workflows that span multiple platforms and datasets, Swanson said.
“There are significant opportunities across your business processes but you have to marry it with the quality of the data and maturity of the full technology stack inclusive of the AI component,” Swanson said.
The safest course of action is for CIOs to insert orchestration, policy and transactional control between agents and systems, said Workato CIO Carter Busse. This will afford IT leaders the power to control “how AI acts across the business,” Busse said.
The narrative arc around agentic AI interoperability or incompatibility might seem familiar to longtime IT pros. One fitting analogy is the network protocol wars from 30 years ago, with IBM, Novell and others vying to standardize solutions before the industry settled on TCP/IP as the internet exploded.
Even so, consolidation among agentic standards will happen more quickly than the decade it took for TCP/IP to become dominant.
“All this [agentic AI] stuff is just in its infancy,” Dayaratna said.
