The phone is still the API
This month the industry settled where agent-to-agent protocols live. This summer Google's agent started phoning pet groomers. Both make sense, and together they say a lot about where integration actually happens.
Two things happened this month that I keep putting next to each other.
On August 17, Google’s A2A protocol joined the Agentic AI Foundation, the Linux Foundation body that already hosts MCP. Over 250 member organizations, every major lab and cloud. There is now one agreed way for an agent to discover another agent, hand it a task, and get a result back. JSON over HTTP, signed agent cards, version negotiation. Proper plumbing.
The same summer, Google’s agentic calling reached everyone in the US. Ask Google to find a pet groomer with a Saturday slot and its agent phones the groomer. It opens by saying it’s an automated system calling on behalf of a user, asks about availability and pricing in plain English, listens to whatever the person behind the counter says, and texts you a summary. Groomers can opt out from their Business Profile, or just tell it to go away.
Look at what the second one is. An agent discovering a counterparty, delegating a task, and getting a result back. It’s the A2A use case, run through a human’s ears.
Nobody at Google thinks a phone call is a good API. They shipped it because the groomer doesn’t have an A2A endpoint. The groomer has a phone, a paper diary, and a person who picks up. That’s the interface. A protocol is worth exactly as many counterparties as speak it, and for most of the businesses you’d actually want an agent to deal with, that number is zero.
None of this is new. Purchase orders went over fax for decades after EDI existed, because the supplier had a fax machine and no EDI. Fintech spent ten years scraping bank websites with a stored password because the bank had a website and no API. Every generation of integration starts by pretending to be a human at whatever interface the other side already has. Voice plus plain language is just the most general version of that trick we’ve ever had.
The next step is already happening, and it’s a little absurd. The same groomers are being sold AI receptionists, and plenty of them will buy one. So within a year, a meaningful share of these calls will be one language model phoning another language model, each running speech-to-text and text-to-speech, both paying for tokens, to exchange something that fits in 200 bytes of JSON.
“Hi, this is an automated assistant calling from Google on behalf of a customer. Do you have any Saturday availability for a medium-sized dog?”
“Thanks for calling Paws & Co! I’m the virtual receptionist. Let me check that for you.”
Two computers, doing impressions of two people, over a phone line. It’s funny until you remember that this is exactly what the fax era looked like from the inside.
Modems solved the same problem in the 1980s. Two machines on a voice line would play a tone, recognize each other, and stop pretending to be voices. A2A’s agent card is that handshake over HTTP. Somebody will build the audio version: an agent that answers the phone, hears the disclosure, and says “if you’re an agent too, here’s an endpoint, let’s do this properly.” If you’re building the receiving side, I’d build that first.
The more useful lesson, though, is about the businesses, not the agents. An agent phoning your business is a symptom. It means the business has no machine-readable surface, so the agent fell back to the one interface you do expose, which is a person. The fix isn’t a better phone bot on your end. It’s the boring stuff: a calendar with real availability, a booking endpoint, a price list in a format that something other than a human can read.
I wrote earlier this year that most digital transformation is renovation. This is the same lesson from the other direction. The businesses that expose a structured surface will get booked by agents quietly, in the background, without anyone picking up. The rest will get a polite robot on line two, asking about Saturday.