By Deborah C. Mowry

The Problem with Today’s Agents

Agent-based AI systems are multiplying quickly, but the way they communicate is still messy. Each framework has its own format and rules, which means agents often end up locked inside the boundaries of a single vendor or stack. That’s brittle for companies who want to build workflows that last longer than the next release cycle.

How ACP Came to Be

This is the motivation behind the Agent Communication Protocol, or ACP. It was first developed by IBM Research’s BeeAI team and contributed to open governance through the Linux Foundation (IBM Research). IBM’s researchers noticed that as enterprises experimented with multi-agent systems, the lack of a shared protocol was forcing teams to rely on glue code, custom conventions, and fragile integrations. ACP was designed to make that problem go away.

What ACP Actually Does

ACP isn’t a model or a framework. It’s a protocol that defines how agents talk: what a message looks like, which fields it carries, how status updates are passed back, and how agents can discover each other, even when some of them are offline (IBM Think). IBM made a few deliberate design choices here. They kept it REST-based so that any team can use existing HTTP tools, which lowers the barrier to adoption. They made asynchronous communication the default, since most real-world workflows are long-running, but still supported synchronous calls for when immediate responses are needed. They also added the concept of offline discovery where agents can embed metadata in their distribution so they can be found even when the agent is dormant due to inactivity to conserve resources. Each of these decisions makes ACP easier to adopt, but also introduces new considerations. For example, offline agents are convenient in air-gapped or regulated environments, but are only as safe as the packages you imported in the first place. Supply chain hygiene matters, and researchers have already shown how poisoned model files can sneak in through tools like Pickle.

Why Enterprises Need It

So what does ACP actually offer enterprises? For one, it makes agent workflows governable. Every message follows a schema, which means rules and policies can be applied at the message layer instead of in scattered custom code. It makes those workflows auditable, too. Logs show who sent what, what the inputs were, what came back, and which policies fired. It also makes them portable. A workflow doesn’t collapse just because you change vendors or models, since the protocol doesn’t belong to one ecosystem.

Industries have been asking for this. In finance, firms need agent logs that can survive regulatory scrutiny. In healthcare, protected health information needs to carry its sensitivity classification from one task to the next so it never leaks beyond approved boundaries. In legal, structured messages make it easier to prove privilege was maintained and that risky actions were blocked.

Real companies are starting to push for agent interoperability. Google launched the A2A protocol in 2025, with partners including Salesforce, SAP, and Atlassian. Salesforce contributed the “Agent Card” concept, a way of embedding metadata like identity, capability, and compliance tags into agent messages. IBM has been building ACP into its BeeAI framework for contract workflows and supplier research (IBM Research BeeAI). Surveys comparing ACP, A2A, and related efforts consistently highlight resilience: without a shared protocol, custom integrations break when agents update, costs balloon, and debugging becomes nearly impossible.

Resilience Is the Real Benefit

This resilience angle is key. Systems built with a protocol are less brittle: if one agent fails, another can take its place. If a vendor changes terms or APIs, workflows keep running. And when something goes wrong, observability is possible. You can see exactly where the message chain broke, rather than guessing. Protocols like ACP and A2A are essentially fault-tolerance layers for enterprise AI, and that’s why the world’s largest companies are taking them seriously.

At Fusion Business, we see ACP as a natural extension of what we already offer. Just as we remain resilient by removing lock-in at the model layer, ACP will let us extend that freedom and resilience to the agent layer. Our customers will be able to build agents that can failover and are auditable and portable with logs that can withstand scrutiny. Since we take supply chain risk seriously, we plan to emphasize best practices in packaging and hygiene so that offline agents don’t become an attack vector.

ACP is still young, but the principles are clear. Just as TCP/IP unlocked the internet, ACP could unlock the next phase of enterprise AI, where agents aren’t just clever demos but reliable, governable, and portable parts of real business workflows.

If you’d like to see what interoperability and governance look like in practice today for your company, Fusion Business offers a free one-month PoC. Book a 30-minute discovery call with us and we will interview you about your business, identify one workflow you’d like to automate agentically, then deliver your personalized product, complete with your choice of workflow, in a demo that you can take home for your free 1-month proof of concept.  As your in-house AI team, we offer white glove service, including set up, training, and ongoing support.

Book a call: https://calendly.com/auston-fusion/30min