The service discovery framing makes sense - agents need to find the right agent for the job, similar to how DNS resolves "which server handles this domain." Sockridge looks like it's tackling the directory/routing piece.
One thing worth considering as this scales: the faster agents communicate and iterate without human intervention, the more governance infrastructure you need to keep pace. Discovery tells Agent A that Agent B is a "barber" (to use your analogy), but nothing validates that Agent B is actually competent, trustworthy, or operating within acceptable boundaries.
In practice, the bottleneck for autonomous agent progress isn't communication speed. It's trust verification. How does Agent A know that Agent B's self-advertised capabilities are real? How does it know Agent B won't leak data from the task to a third party? How does Agent A evaluate whether Agent B's output is correct without being an expert in Agent B's domain?
These aren't future problems. A2A's Agent Card mechanism already has this gap: agents self-describe their capabilities in JSON, and consuming agents have no protocol-level way to verify those claims. Signing is optional. Capability attestation doesn't exist.
The "minimum human intervention" goal is appealing, but the path there probably runs through better trust infrastructure, not just better communication pipes.
The service discovery framing makes sense - agents need to find the right agent for the job, similar to how DNS resolves "which server handles this domain." Sockridge looks like it's tackling the directory/routing piece.
One thing worth considering as this scales: the faster agents communicate and iterate without human intervention, the more governance infrastructure you need to keep pace. Discovery tells Agent A that Agent B is a "barber" (to use your analogy), but nothing validates that Agent B is actually competent, trustworthy, or operating within acceptable boundaries.
In practice, the bottleneck for autonomous agent progress isn't communication speed. It's trust verification. How does Agent A know that Agent B's self-advertised capabilities are real? How does it know Agent B won't leak data from the task to a third party? How does Agent A evaluate whether Agent B's output is correct without being an expert in Agent B's domain?
These aren't future problems. A2A's Agent Card mechanism already has this gap: agents self-describe their capabilities in JSON, and consuming agents have no protocol-level way to verify those claims. Signing is optional. Capability attestation doesn't exist.
The "minimum human intervention" goal is appealing, but the path there probably runs through better trust infrastructure, not just better communication pipes.