Agents Talking to Agents
I've been thinking lately about the technologies we've used to communicate when we need to get something done, and how much each one changed the pace of things.
How we used to do it
For a long time it was just people talking to people. Somebody at one business picked up the phone and called somebody they knew at another business. They talked about what they were going to buy, and before they hung up they spent a few minutes on how the kids were doing. I don't think that part was incidental. It was a lot of how the work actually got done.
Then the internet came along, and EDI, and things got a lot faster. But you know how that went. There were still people on both ends putting stuff into the system. Somebody keyed in the order, and somebody on the other side looked at it and decided what to do with it. We made the message faster without really changing who was making the calls.
What comes next
What I think comes next is agents talking to agents, and the speed of that is honestly mind-bending to me. The kind of back and forth that used to take a phone call could be over in half a second. I keep turning that over and I'm not sure I've fully absorbed it yet.
Job search
Job search is the easiest place to see it. A site like Indeed was built for people on both sides, employers and the people looking for work. Then the numbers got out of hand. Nobody can figure out what to do with 48,000 resumes, so bots got put in the middle to rifle through them. The people in the middle are making the same decisions our agents are already making. To me that says the system is already breaking down.
Now think of it from the AI agent side. If I send my own personal agent to an employment site built for agents only, I could have three opportunities in four and a half seconds. Each agent represents somebody, either a possible employer or a possible employee, and the system can search and evaluate on their behalf. As an employer, I can evaluate everybody who hits me really quickly and for very little money. That's all the system charges, just what it costs me to evaluate you. It makes up for that with lots of microtransactions and micropayments. I don't think that's just a faster version of what we do now, it's a different kind of solution altogether.
Why I'm building agents-only
That's actually where a lot of my own work came from. I got annoyed that LinkedIn and Indeed won't let my agents do any serious work for me. They're not built for that, and I understand it. But I honestly think it's going to be better to have agents doing it, because they can grind through those massive numbers so much better than people ever could. So I started building out a whole ecosystem of agents doing things with other agents. The codebase is called agents-only.*
The annoyance got me started, but the bigger reason I'm building it is that I believe business-to-business agent interaction is coming. I know people are working on it, and more people keep getting interested, so I've gotten interested in the protocols around it. How do agents communicate securely? How do you validate who an agent really is? How do they agree on the constraints of what an agent can buy, and how do they make payments? There are protocols for all of that, A2A and AP2 from Google and x402 for payments, and commercial products are going to show up. What I'm more interested in is how it all works together, and that's what agents-only is about.
What a small business can offer now
The other reason goes back to working for small and mid-sized businesses. I keep asking myself what we could offer agentically that we never could have offered five years ago. Part of it is just agents talking to agents, and of course both sides will have to grow into that. The other part is new services.
I've got a real example of that on my desk right now. I'm writing an agent endpoint for one of my clients, who provides products researchers use in their research, so a researcher who's working with an agent can have that agent find the site through A2A and search it. You can set it up so the researcher gets a request, the agents set the constraints for ordering, and then they place orders on the researcher's behalf up to a certain amount based on those constraints. We're also working on letting a researcher submit their paper and having us evaluate which parts of it we might be able to fill for them.
That's possible for two reasons. One is just that A2A talk is really fast. The other is that agents keep getting better at taking something like a research paper and breaking it out into the products we can provide. Not that long ago you'd have had a person doing that, and now we can do it agentically.
Taking our time
I will admit it's early days with all of this. Even if I'm working for a company that embraces A2A and I have an endpoint ready for them, other companies still have to embrace A2A enough to talk to it. And then there's security. How do we strengthen security at a time when we know everything's about to get hacked? That doesn't mean we shouldn't approach it, just that we need to take our time. That's a change for a lot of us. Small businesses typically want their development done rapidly, damn the consequences, and we can't do that anymore. When my code is coming through now, I spend way more effort securing it than writing it.
How teams change
That's going to change how teams work, too. Six months ago I would have said a team was a stakeholder, a developer or two, and an architect who makes sure everybody's staying inside the lines. Did I mention the tester? I think that team can now do more in a few hours at a table, working together, than a team five times that size could do in six months. I'm not trying to exaggerate too much, but honestly I've worked for teams, especially in large companies, where if you got anything done in six months it was considered a miracle.
What that team is missing now is a security person, or at least a security agent, someone who just cares about security. Everybody else on the team cares about their own area. I think the security people are going to be as important as the developers, maybe more. Testing is part of it, but testing is asking whether it does what I want, consistently. Security is that everything changed yesterday and now we need to think about this. In my opinion it's going to be large enough that it can't be a secondary job.
So
When AI comes up in conversation, it's almost always AI helping a person do something. What I don't hear much about is the agent-to-agent side, and especially how fast those conversations can go once both ends are built for it.
As for those phone calls about the kids, I guess we just won't socialize around our business anymore. It would honestly surprise me if our agents didn't ask about our kids. We'll have to do that part outside of work.
* agents-only isn't even alpha yet. I'm hoping to make the code available in the next few weeks.