Herdr's Agent-to-Agent is SHOCKING!!!! 😱🤯💥 transcript

Greetings everyone. I'm Furu here. Today I want to show how agent to agent communication works within Herder. This will be a deeper presentation to the one we did a little bit earlier emphasizing some different scenarios and just why this is an amazing capability to have. With that said, we're going to jump right in to our different scenarios. Here we have agent A, which is a Claude session. And I'm using agent here a little bit loosely.

Each tab is considered an agent. Agent B is another Claude session and I'll go ahead and clear this. Agent C is not Claude. This is a scenario that we all too may be familiar with where you have, in this case, Codex and throw in Antigravity, and throw in Hermes, and throw in Pi, throw in Open Court. There are many scenarios where you may want to have your agents talking to each other across not just providers, but across sessions and across different instances.

Extremely powerful. And Herder provides a very compelling primitive to help us do that. It's not the only way to do it, but it's one of the reasons why Herder is becoming more and more popular. Now, let's start with our very first scenario here, which is the agent in Claude. If Claude is all you use, it's a it's a no-brainer. I think we know exactly how Claude has implemented uh uh agent communication within Claude itself.

If you go into Claude, you do {forward slash} agents, you'll see a list of your library of agents. So, in my case, uh these are some agents in here. Uh a spin content writer agent, a back-end agent, a front-end agent, a QA agent, resource agent, you name it. And what is an agent? For the sake of definition, an agent in this case is essentially uh context, instructions, and a model. Combine those, you have an agent. Okay, and I know people use agents to mean different things.

Again, context being uh what space is agent going to work with? What model is it going to use? In this case, Sonnet. What are the instructions specific to the way this agent will behave? In this case, how does it write content about Spain? That is an agent definition. And so, with this in mind, again, just because uh this is a refresher, I think for most people, I can come into this particular session and do something like write uh 100 words about Spain using this agent.

And because this is Claude, and the agents was registered in Claude, an instruction like this executes pretty well. It would take that, invoke that agent. You can see the agent uh showing up down here. That article would be written. It's disappeared. And you'll get your results back. That is no-brainers. Claude has really, really done this well. And because the way this works is this name now almost becomes an address to that agent, and it becomes an addressable point.

So, if I want, instead of a 100 words, just to illuminate on this a little bit, and then we'll move on, I wanted uh 500 words about Spain. So, it will spin up that specialized agent. And if you want, you can actually go into the specialized agent like I did here. This is where it's writing, and I just simply went down and tap into Spain writer agent. Uh you can go back and forth, back into your main agent, or you can go down back into your specialized agent.

Very, very simple. You can have dozens of these agents, your research agent, email agent, real estate agent, business agent, vacation agent, you name it, you have all of these available for you to work. So, a very, very, very compelling. Now, the reason why this is good is because if I have a session here and this has successfully written about Spain, >> [snorts] >> this we are not limited to the session. We are not limited to a session to invoke this.

I can go to a new session, let's go over to Claude B and paste the same thing, this will work. It will know what Spain is, what this agent is. And you open up any new tab and put in this prompt, it will work. All fine and good, but there is a caveat. And that caveat is if all you do is within Claude. But here is a scenario for most people. I use Claude, sometimes I'm out of tokens or for whatever reason, I want to get the perspective of Codex.

Will the same prompt like this work? And I think guys, I'm sure you all know what the answer is. Would the same prompt work? Absolutely, I don't expect it to work, but let's try. You never know. So, I'll go into my Codex session, paste the same prompt, write a 500 words write 500 words about Spain using agent this. All right? And let's see. Uh it's using GPT 5.5, which is interesting. And let's do the same with anti-gravity.

And this access denied, there is an issue with that. One, which is fine. We'll stay away from anti-gravity for today. And ChatGPT having some issues, so let's do models. Let's do the default models, do that. Try one more time. Write that. And my expectation is that it shouldn't it shouldn't work. Okay? Um it shouldn't work. And the reason why it doesn't work is uh it's looking to use open rig in there I don't have open rig installed.

Um well, I did, but this video is not about open rig. Uh bottom line is it wouldn't work. Uh just ignore the open rig for now. So, let me All right. So, we've taken care of the open rig piece, um but you can tell it says I couldn't reach any search agent and and this is where Codex behaves in his own way sometimes. Somehow it somehow wrote this other code and doesn't claim to having written by an agent. I'm I don't appreciate the behavior, but that is Codex.

Uh coming back to Claude, we now see a scenario where if all we used was Claude, we can create these agents, use them within Claude, and reference them easily. But, what if what if I have desktop open and all I wanted to do was to pass a message to this other tab here. So, agent A Claude talking to agent B Claude to tell them hello. Let's see we can do that. So, can you say hello to agent B Claude uh using her just like that.

Now, I have this and let's just accept the commands here. Claude has opened up the pane. It's done that. Uh if I go over here, you can see hello uh from uh Claude A says hi. All right. So, now I can say 100 words about Spain. So I can now from within A pass that message. If you come over here, the message was just received and is invoking and is going to write 100 words about Spain. This is what Herder allows us to do.

This now becomes interesting. If you're like me, you can now see that this is becoming interesting. Why? Because if I wanted to do this from agent A, not just to agent B that is in Cloud, but I want to do it to agent C that is in Codex. Remember we were struggling with this particular agent in the behavior of that agent. Uh let's try it. So let's do um agent um C Codex. Just bear with me. So agent C is what? Codex. And when I say, "Can you say hello to agent C Codex?" Um and this is saying sending now.

In theory, I wouldn't have to always bypass permission here. And now agent C, "Can you write of what about Spain?" Remember agent C wasn't wasn't Cloud. We couldn't do it if we just went through the Cloud mechanism. Now if I go into agent C, do you see that? It took It received the instruction. That instruction was received. And it wrote about Spain. Now you might say, "But Fru, like there's too many Spains going on.

Like I You pissed at this." No, I did not. Let's go. Give me your favorite country. Uh not Spain. Let's do Kenya. All right. I didn't write about Kenya. Now quickly switch over here. Did you see that? Hello again. Now I'm in Codex. I received your message about Kenya. I will write about Kenya in East Africa. Isn't that something? Isn't that something? Now, all of your agents becomes addressable points that you can talk to in Herder.

That is amazing. Now, this is anti-gravity. This is a Hermes. You might say, "But, well, could we do it with Hermes?" Yes, you can. You can, actually. It's It's that It's that fascinating. All right, let's not do Kenya. Let's do Mexico. Let's do Mexico. All right. And where do we want to write the agent to or from? We do F Hermes. Right. So, now I'm saying, "Can you say hello to" I'm I'm here in agent A and I'm sending a message to agent B to do some stuff.

And you can tell that when it does its work, it comes back and acknowledges what it did. You You could maybe want that or not want that. So, um uh Now, let's go to Hermes. All right. Hermes did receive the message. Uh and it's asking for my key. So, we're running into some tokens issues there. But, it doesn't take away from the fact that Hermes did receive the message. All right. So, just like this, we now have a plane of interoperability where within Cloud, you can send message to these addressable endpoints.

They will receive that, work on that, and then bring the message back. And I think this is extremely compelling. There are other ways to do this. There are other tools that can do this. But, the fact that Herder makes it so easy. Now, if agent A, agent B, agent C is not as compelling, imagine this is your front end. This is your back end. This is your security. This is your QA. This is your um the website for an app you're building.

Right now, you can have QA using Codex. You can have graphics design using Claude um on um Hable. And then you can have this running on Opus 5.5. This is how Her then now becomes an extremely compelling mechanism. And the reason for for this is if you notice, Her has this agent prompt that you simply uh can reference the prompt. So, every tab gets an ID. Uh and then I'm not going to the technical details here. I wouldn't claim to know how Her is implemented.

But just from what I'm reading here, every uh tab gets an ID with a name. So, once you have that ID, uh you can essentially uh then uh uh message uh into that uh into that endpoint if we um if we um if we want. And this is really really cool. And now because Her Her Her uses Tmux behind the scenes and is long-running, you can essentially just have this as your agent topology on a Mac mini running. And I've done some videos on how you can have this remote using something like a Mosh or whatever means.

And now your agents are just talking and passing messages to each other. So, let me do this. Check what all the message what all the agents in the open tabs have done. Give me a summary. Short. All right. So, just like that. I'm simply saying, "Hey, for you, Claude, go check on all these other ones. See what they've done. And give me a summary here." It keeps asking for permission. And so, you can tell it's gone. It's seen the message, what they did.

Gemini did not behave and is giving me a report. All right. Now, this now can become your chief of staff agent, being delegated to all these other agents. And I think this is extremely extremely uh compelling with that mechanism that um Herder provides. Now, Herder also has the remote uh connection. So, this is on the MacBook Air. If I have my Mac mini, imagine now delegating the work to your Mac mini from your Mac Air with one chief of staff, and Herder is doing the messaging.

I I think that if you're not using Herder just for this alone, I I don't know what is. I I think it's extremely compelling. There are solutions to this that I've come across. Let me touch on a few. So, OpenRik is an interesting solution that seems to do something to the effect of um played with it, done some demos. Um a a very interesting project. Let's see on GitHub how many stars do they have. About 500 stars. Respectable.

Uh project last commit just about 11 hours ago, so still active. And it gives you a very opinionated way to create addressable endpoints to your agents using this methodology called Riks. And then, you can address your agents regardless of what tool you're using. So, Claude, Gemini, Codex, Pi, Hermes, you name it. I thought that was very very compelling. Uh you can do that also within uh within Hermes or within Herder.

Uh so, an interesting one. And ultimately, the the end goal is being able to have Gemini, Codex Codex or Claude talking, passing messages, learning from each other, delegating work, and you can arbitrage and use the best model for the job in real time uh without being tied to one model provider. And I think this is a very compelling approach. So, hopefully this helps uh gives you some some ideas on how to to use it um either the the regular agents that Cloud provides very well or just going in with um Herder and having Herder do the messaging for you, which it does also very very interestingly.

And the fact that it maintains all your conversations within uh the session as such comes with its own benefits because sometimes I I want to go back and read the conversations for each particular agents. And if you're not happy with that, you can always you can always at some point say, "Hey, save to disk." All right, you can have a a hook or something to say save this to disk. So, you can write your your chat history into disk for for more longer-term persistence.

So, hopefully this is helpful. As always, this is Fru. Thank you for watching. I'll see you in the next demo. >> [music]