Agent mesh

As I wrote in earlier posts e.g., When to multi-agent?, I believe agents should work in a network, not a hierarchy, similar to how the internet evolved. I heard the phrase agent mesh recently, which covers that idea well.

A colleague pointed me to this post by the Confluent team the other day. Think of it as enterprise application integration for agents, or if you prefer, enterprise agent integration. At a conceptual level, I subscribe to this view, as I expect autonomous agents to sit at the heart of future business process automation.

Where does Kafka come in?

The Confluent article lies at the intersection of three challenges.

1. Agent communication inside the organization. I doubt you would expose Kafka directly for inter-organizational traffic, although an API façade could work.
2. One-to-many or many-to-many patterns. For one-to-one or many-to-one, you can probably use agent-to-agent protocols without an event broker.
3. High-volume event-based workloads that need retention and replay—far beyond what a simple message broker offers.

Questions worth asking

When I think about a preferred agent mesh architecture, a few questions immediately pop up.

• Will most A2A use cases be intra- or inter-organization?
• What patterns will "embedded" agents like Slack or SAP support?
• How many scenarios truly need one-to-many or many-to-many communication?
• What volume of interactions should we expect?
• How to address all other non-functional requirements: security, monitoring, observability, and so on?

I hope the answer to the first question is "inter-organization", fostering tighter collaboration between teams and organizations. Intra-organization use cases can be more common, but to a good degree can be handled (more deterministically) by MCP as well. When it comes to one-to-many or many-to-many, I do not expect this to be a common pattern, although persisting events for analytics is always valuable.

User-initiated interactions from employees will likely be low volume. End-user interactions could hit mid to high volume. Machine-triggered flows—think EDI-style integrations—could be extremely high.

Crawl, walk, run

The Confluent post is thought-provoking, but we need to crawl before we try to sprint. Today, very few agents are live in production, and even fewer rely on agent-to-agent protocols.

Let's find concrete use cases, put them in production, scale them, and only then decide how to avoid the quadratic explosion of pairwise interactions. By then the protocols, the technology landscape, and our own understanding will have evolved.