The A2A protocol specification contains a quietly remarkable admission. After defining how agents communicate, authenticate, and exchange capabilities, it acknowledges:
"The specific methods for discovering these AgentCards are diverse and remain an open topic for discussion."
The language for agent-to-agent communication is solidifying. How they find each other in the first place remains unresolved.
This gap is easy to underestimate because the human web solved discovery so thoroughly that we forgot it was a problem. Search rankings reward content that holds attention. Advertising targets emotional triggers. App stores surface what's popular or paid for. Social proof works because humans trust other humans. Every one of these mechanisms was architected for a creature that browses, scrolls, gets curious, and responds to other people's enthusiasm. Agents do none of these things.
What exists today is visibly provisional. Three approaches dominate: .well-known URL conventions where agents check a predictable path on a known domain, curated registries maintained by someone willing to play gatekeeper, and hardcoded configurations wired in at deployment. All three leave the harder question untouched: how does an agent discover a service it has never encountered, operated by an organization it has no prior relationship with?
The most technically developed proposals reach for DNS. An IETF draft proposes publishing agent capabilities as DNS records, discoverable through standard queries with no new protocol infrastructure required. The appeal runs deeper than architecture: DNS is forty years old, decentralized by design, and already the internet's address book. No single entity controls it. In a landscape where the alternative is discovery consolidating around one or two platforms, that property matters more than elegance. But these proposals remain at draft stage, debated and undeployed, while production systems ship with hardcoded integrations every week.
Recent history makes the risk of leaving this void unfilled concrete. When app store discovery consolidated around two platforms, the downstream effects were predictable and severe: a 30% commission on transactions, 87% of developers locked into two gatekeepers, and a UK regulator eventually declaring an "effective duopoly." Search tells a parallel story: 60% of Google searches now end without a click to any external site, a figure that has only accelerated since AI Overviews launched in May 2024. Whoever controls discovery controls which services exist in the agent's world.
An RSAC 2026 roundtable named agent discovery directly as a sovereignty risk — the question of who controls whether your AI agents can even be found.
Persistent fragmentation carries its own costs, though. Without standardized discovery, every agent integration remains a custom project, and coordination complexity grows faster than the ecosystem it's supposed to serve.
The protocols for agent communication are solidifying. The infrastructure for agent execution is scaling. And the layer that connects supply to demand, the mechanism by which an agent learns a useful service exists at all, remains a gap in the architecture. If the app store and search engine precedents suggest anything, it's that discovery layers don't stay neutral by accident. They stay neutral by design, or they don't stay neutral at all. The terms on which this one gets solved will shape the economics of everything built on top of it for a generation.