MCP is the only one of the three that looks load-bearing in production right now. A2A and ACP feel like specs for a multi-agent world that hasn’t shown up. Discovery and agent cards start mattering once agents from different vendors have to negotiate with each other, and most setups are still one orchestrator calling tools. Are you seeing A2A used anywhere real yet?
Data point for the load-bearing question. I sweep the public MCP registry with an open tool and ran it twice: 16,548 servers on 14 July, 22,659 on 19 August. Roughly 175 new a day. Whatever else is true, that one is not a spec waiting for adoption. The more interesting part is what moved underneath it. In July, servers shipping as an installable package were the majority at 50.4 percent, hosted remote endpoints at 42.6. Five weeks later that inverted: 45.8 package, 47.3 remote. That is the discovery and trust problem arriving through the back door. Not agents negotiating with agents, but tools you call over someone else's URL, with no version to pin and no changelog that reaches your dependency scanner. One caveat on reading registry counts as production reality: on a seeded random sample of npm and stdio servers I am seeing an inclusion rate near 59 percent, against 67 percent on a hand-curated list. Interim, n=61. Most of the failures are servers that never start.
Correcting my own numbers above, same day. The 19 August census I cited could not be reproduced from committed scripts, so I have withdrawn it rather than leave it standing. The verified re-run, 14 July to 22 August, 39 days: 16,548 to 24,135 servers, about 195 a day. Package-only 50.4 to 43.6 percent, remote-only 42.6 to 49.7. So remote did not edge ahead, it crossed over, and my figure understated it. The behavioral number was worse than I said too: the full seeded random sample of 400 came in at 195 included, 48.8 percent, not the 59 percent I quoted off a 61-server slice. Handshake failures 150 against 53 credential requirements, so the shape of the finding holds. One thing that cuts the other way and is worth saying: across the 2,766 tools those servers exposed there were zero fatal schema violations. The schemas are fine. Reachability is the problem. Every figure in this reply regenerates from a committed script. The ones above did not, which is exactly why they are gone.
Clear breakdown of a confusing space. One angle I would add from the self-hosting side, once you are wiring agents together with these protocols the bottleneck quietly moves to how many models you can actualy serve at once on your own hardware. A chatty multi-agent setup can fan out into a lot of concurrent inference, and that is where a second GPU and batched serving earn their keep versus a single card.
Great architectural breakdown. Standardizing agent-to-agent communication protocols feels like the early days of establishing HTTP for the web. Getting this right is going to be the difference between fragmented, siloed automation and a truly interoperable AI ecosystem.
MCP is the only one of the three that looks load-bearing in production right now. A2A and ACP feel like specs for a multi-agent world that hasn’t shown up. Discovery and agent cards start mattering once agents from different vendors have to negotiate with each other, and most setups are still one orchestrator calling tools. Are you seeing A2A used anywhere real yet?
Data point for the load-bearing question. I sweep the public MCP registry with an open tool and ran it twice: 16,548 servers on 14 July, 22,659 on 19 August. Roughly 175 new a day. Whatever else is true, that one is not a spec waiting for adoption. The more interesting part is what moved underneath it. In July, servers shipping as an installable package were the majority at 50.4 percent, hosted remote endpoints at 42.6. Five weeks later that inverted: 45.8 package, 47.3 remote. That is the discovery and trust problem arriving through the back door. Not agents negotiating with agents, but tools you call over someone else's URL, with no version to pin and no changelog that reaches your dependency scanner. One caveat on reading registry counts as production reality: on a seeded random sample of npm and stdio servers I am seeing an inclusion rate near 59 percent, against 67 percent on a hand-curated list. Interim, n=61. Most of the failures are servers that never start.
Correcting my own numbers above, same day. The 19 August census I cited could not be reproduced from committed scripts, so I have withdrawn it rather than leave it standing. The verified re-run, 14 July to 22 August, 39 days: 16,548 to 24,135 servers, about 195 a day. Package-only 50.4 to 43.6 percent, remote-only 42.6 to 49.7. So remote did not edge ahead, it crossed over, and my figure understated it. The behavioral number was worse than I said too: the full seeded random sample of 400 came in at 195 included, 48.8 percent, not the 59 percent I quoted off a 61-server slice. Handshake failures 150 against 53 credential requirements, so the shape of the finding holds. One thing that cuts the other way and is worth saying: across the 2,766 tools those servers exposed there were zero fatal schema violations. The schemas are fine. Reachability is the problem. Every figure in this reply regenerates from a committed script. The ones above did not, which is exactly why they are gone.
現状のAIエージェントの乱立する状況を無理やり3つに分類してパッケージ化しようとした結果、技術的な定義が崩壊している
ACP was merged into A2A a year ago - unless I’m missing something I don’t see why they’re being presented as alternate options
Clear breakdown of a confusing space. One angle I would add from the self-hosting side, once you are wiring agents together with these protocols the bottleneck quietly moves to how many models you can actualy serve at once on your own hardware. A chatty multi-agent setup can fan out into a lot of concurrent inference, and that is where a second GPU and batched serving earn their keep versus a single card.
Great architectural breakdown. Standardizing agent-to-agent communication protocols feels like the early days of establishing HTTP for the web. Getting this right is going to be the difference between fragmented, siloed automation and a truly interoperable AI ecosystem.