Latest / The AI Ecosystem Podcast / When Agentic Collaboration Crosses Company Lines: Marketplaces and Agent Cards

Transcript

When Agentic Collaboration Crosses Company Lines: Marketplaces and Agent Cards · The AI Ecosystem Podcast

• Bob Moore: Hey everybody, welcome again to the AI Ecosystem Podcast. Never have the words AI and ecosystem been showing up this frequently and this close together. Our listenership continues to build. I was pleasantly surprised with the random emails for swag that we got last time around. And people are listening. I'm I'm being way too self deprecating. This this podcast has legs. I'm Bob Moore, CEO, co founder of CrossMedium here with Lindsay DeFalco, chief. product officer, we've got a nice big agenda today, but we want to start with a really cool piece of work that we've been doing around hyperscaler marketplaces that we announced yesterday. Lindsay, do you want to give a little, a little rundown on what that is? Lindsey DeFalco: Yeah. Hello everyone. all right. So yeah, we're actually going to talk about ecosystems and AI together, which we don't do on every podcast. so like Bob mentioned, we announced yesterday that we were integrating with AWS's lead enrichment product that they launched. and we've talked about the marketplace angle before. As a reminder, this is Very, very interesting right now because these transactual marketplaces that have existed for years and years and years are just going to become more important as agentic transactions become reality because they're essentially these places that already provide all of the guardrails around which to transact. And so it's a fascinating space in general as the what agents can actually do is evolving, and as we move into a world where you know, agents may actually be the ones like procuring and purchasing software. and so for that reason we're we're i excited about what is coming there. Bob, I don't know if you want to say anything else about like the general kind of marketplace landscape before I get into what we're what we actually did there. Bob Moore: Yeah, I I do think there you don't have to squint that hard to see the connection between this stuff and and AI. But before we do that squinting, let me just spend two seconds on for those who are new to hyperscalar marketplaces and what they do and why they matter. these things have existed for over a decade. Amazon AWS is is the pioneer in this area. And it started out as a really simple concept, which is a lot of companies Sell stuff that actually has to be stood up and run on AWS servers and that consume AWS storage. And it just made logical sense that maybe AWS should effectively be reselling that stuff on behalf of the companies that make them. Because back in the day, you'd stand up these containers that might include a piece of software that a company has sold to you or you bought a license to. Well, that has evolved over time to a point where It now is this enormous line of business and source of transactions for Amazon, where you can buy virtually any software product from Amazon. And Amazon's marketplace sells products and services now that may or may not directly correlate to AWS consumption happening at all. But if you want to buy, for example, if you want to buy Crossbeam and you're already an existing AWS customer, And you'd rather not have another vendor and another bill to pay and another PO to cut and all this other stuff. You can just buy it through Amazon on your existing Amazon account, and Amazon keeps a very small percentage of that, like a credit card company would, and passes the rest of the revenue back to us. they've made this extra compelling by allowing money that's committed. To be spent with Amazon. So if you're an Amazon customer, you often make these spend commits in order to access a discount. So if you're committed to spending $10 million a year with Amazon and you only spend nine, you got a million dollars left, Amazon will give you relief on that annual spend commit if you buy stuff through marketplace with those dollars. So spend commits become a really big motivation. So there's like security, legal, procurement, and economic. Reasons why buyers want to buy through these marketplaces and sellers want to sell. And Amazon has has done a fantastic job. Microsoft with their Azure Marketplace has a really compelling business there too. GCP is in this business as well. you know, we know that there are there are other emergent ones from providers in more vertical SaaS areas where marketplaces are starting to bubble up. so what does all this have to do with AI? Well, If you envision a future where agents are buying stuff, agents need a place to transact. And while my consumer agent, and we'll talk about Muse and Instinct in a in a bit, my consumer agent, I can just hand them a credit card and they can go buy me a new Wi-Fi router on Amazon and it's is it's easy peasy. When you're buying stuff as a business, it's not that easy peasy. You're very often entering into a contract with certain obligations and legal protections. You need to Actually, transfer money through a bank wire and you need purchase orders, and there's tax implications that are are super complicated, and there's process wrapped around buying stuff. So if everybody in big enterprises just handed all their agents a bunch of credit cards, all hell would break loose. Agents need to know where they can go to actually transact and buy stuff without falling outside of all these policies and processes for procurement. And the marketplaces turn out to be a fantastic answer for that, that kind of have solved that problem already almost by accident. or maybe we should give AWS Marketplace credit a decade ago for knowing that this was coming. But at the end of the day, allowing your agent to spend on AWS Marketplace against your AWS spend commit is a much easier proposition than handing it a credit card and telling it to go buy licenses for a new software product. so If we believe in A to A commerce and A to A procurement existing in the future, then these marketplaces are very likely the rails on which a lot of that will happen. So it it ends up mattering not just because they're capturing a lot of the GDP of the information economy, but because the future means by which you transact. Like there, there's market making going on here that has has big implications in the AI era. So we're leaning into it because we believe that. And historically we haven't spent a lot of time in in marketplaces, but some changes recently in what AWS offers from an API standpoint have allowed us to start playing ball there in a big way. And we released at least a first foray into that recently. So maybe Lindsay, do want to with that with that precursor out of the way, Lindsey DeFalco: Yeah. Yeah. Bob Moore: do you want to talk about what the thing actually is we just released? Lindsey DeFalco: Yeah, so if we you know, Bob's going into detail about this this future where the agents are transacting, but in the meantime, when there's still humans and and reps involved on both sides, there is a need for as much information to pass between Amazon and their sellers, so the people that are actually selling on Amazon, as possible. Because obviously AWS wants their partners to close as much business as possible on AWS. and so what they've been doing recently that I think is very cool and very smart is opening up those floodgates a little bit. So they introduced lead enrichment functionality, which allows one of their partners to submit a set of leads into AWS and then it will give them back data that historically you had to actually talk to a human to get from from AWS. So it's things like what solutions are they a good fit for, and engagement scores and other things that are generally like only known by somebody who works at Amazon. You have to talk to somebody who works at Amazon to get this information from them. so this starts to get really interesting. And if you think about like all of these companies that own these transactional marketplaces and what they need to do right now, really they need to figure out how to handle the the volume of Transactions that are coming through and use AI to actually route the correct ones to humans and then have agents help in the ones where humans aren't helping. and so, you know, again, you look at what AWS is doing, they launched an MCP server. and they have one of those for for Partner Central, and so it allows you to actually get some of this information programmatically and have your own agents actually get this information programmatically as well. So if you take all of that together, which is them opening up the gates on information that you used to need to talk to a a person to get, and then also making all of that available for agents and programmatically, then it starts to become speaking of this squinting that Bob was saying, maybe you're squinting a little bit less now. because it becomes pretty obvious how all of that can be useful with you know, accompanying a product like Crossbeam. So all of a sudden we now have access to more information that from directly from AWS that can help you prioritize which accounts to work and whether or not you should work them with AWS or elsewhere. And so what we're doing is we're actually ingesting that. lead enrichment data. And then we're allowing you to enrich your accounts in Crossbeam with that. And so it's giving you this AWS specific Bob Moore: Excellent. Lindsey DeFalco: information directly in Crossbeam where you're trying to figure out where to actually do this transaction with the goal of routing more of those transactions back to AWS and having the content that you know, once you're actually talking to an AWS rep over there, you have richer content because you also have all of the content from Bob Moore: And from Lindsey DeFalco: from your ecosystem and cross beam alongside that that AWS ecosystem. I think that's like functionality wise it in a nutshell. anything you want to add there, Bob? Bob Moore: I just say that the the secular trend here that is coming forth is these marketplaces are moving up the funnel. So the marketplaces used to really, really feel like not a real factor until the minute you were ready to transact. I can't tell you how many deals we've run through Amazon Marketplace the day the deal closed was the first time AWS Marketplace was mentioned. Like we got all the way Lindsey DeFalco: Mm. Bob Moore: through people. You know, we we worked the deal, our sales reps pursued it. We, you know, we won the deal, we got through the RFP, we proved the concepts out, we did the legal negotiations, and then at the last minute, we get the question of, hey, can we transact through AWS marketplace? Because we have some committed spend left over or something. And then we end up we end up doing it. what I think has always been the dream here is that, well. Sometimes AWS's own sales reps are helping you win those deals. Sometimes AWS itself is maybe sourcing the deals for you. These are all things that kind of push AWS's involvement or the marketplace's involvement up the funnel. and obviously that's good for everybody because it helps the sellers win deals and it helps AWS drive like lock in the fact that it's going to transact through marketplace much earlier on. But frankly, those historical cases have been a little a little absent market pool. Like There's only so many leads AWS can generate for people and they need to be fair and thoughtful and take care of their own customers in the right ways. And then there's only so much motivation for companies to register leads or involve AWS early on in the buying cycle. because it's not entirely clear how much it influences outcomes that early on, and it adds just another additional step in the most fragile parts of the buying and selling process. So what this is doing. Which is really smart, is providing a way at scale for Amazon to provide intelligence and inform and direct the way that your deal is is executed without needing to necessarily bother the buyer just yet. and without you know, violating their trust in any way either. It just is a way to make these these very obvious places where there can be collaboration be identified significantly sooner. So that whether you are actually engaging a rep from AWS who's going to help you. Or you're simply bringing up the right topics on the call with the right buyers to help steer a deal in a way that might close faster or close for a larger amount or have a different scoping on it. it's just going to it's going to help. And what that means is that the marketplace part of the conversation starts happening more toward the initial engagement with the prospect as opposed to happening at the very last minute when you're figuring out the the legal and procurement stuff. And that's good for Amazon. And I think it frankly, I think it's good for buyers and it's good for sellers, but This is this is a big this is a big moment, I I think for unlocking that, not just in the world of partnerships, but in the world of sales, which which is a big breakthrough. Lindsey DeFalco: Yeah. And I think it's for me it resonates so much because it is really highlighting the power of data exchange between companies. So exchanging more data ends up making the human interactions higher quality as opposed as opposed to replacing human interactions. So, you know, in a world where you couldn't get on the phone with an AWS rep a certain at a certain account, now you have more information that helps you qualify the accounts that you're requesting in a way that means that your interactions with that AWS rep are gonna be higher quality with more more information in them. So I think with all the the discourse around creating an agent to do XYZ replacing a human, I think in a lot of cases the answer is it's making the human do something different or making them more focused on something different. And this feels really similar where like by opening up this data, they're actually making it so that their reps are spending time on the like better, stronger, higher quality deals and going in with more more information from their their customers. So yeah, fully agree all around. It's like a great example of why CrossMeam exists, because there's power in exchanging data between companies. Bob Moore: Yeah, really cool. and it's like all getting back to the AI topic, right? It's all the pieces falling in place to bring that entire value story into a world where there's agents participating in these economies. and in the same way that, you know, the the marketplaces may enable agents to transact, our goal at CrossFeam here is to enable agents to collaborate, have the information they're supposed to have and that they can benefit from. And make smarter decisions and and take better actions with that information. I think this is this is a very interesting kind of fusion of those two concepts, both of which are very forward-looking in nature, but also as a side effect provide tremendous amounts of value for the extremely human orchestrated world that we live in right now. Because the the the buyers and sellers that are transacting on the marketplace for the traditional reasons right now also benefit from this stuff for the same exact reasons. and that's what we call a you know, have your cake and eat it too moment. could have come Lindsey DeFalco: Yeah. Bob Moore: up with a way better Bob's bingo card metaphor there. but I know. I almost said a triple Lindy, which Lindsey DeFalco: Yeah, it's it was a disappointing metaphor. Bob Moore: is you know, I like to bring out triple Lindy once in a while, but it's not really three things. I could probably make up a third if if I was pressed. Lindsey DeFalco: Yeah, and it's it it's it's also, you know, I mentioned that they have an MCP server and Crossbeam has an MCP server. there's this whole world of value that is unlocked there as well. And we haven't really talked about this, you know, from a marketing standpoint or in in our social media specific to AWS yet. But what's also really powerful about the fact that they're making their MCP server able to access all this data is the the scope of what you're able to do in you know a claw or whatever your client Bob Moore: Mm. Yeah. Lindsey DeFalco: is, it just it blows up. You can put together an account plan for you know, an account that you want to target to have transact on AWS and pull in the entirety of the relevant information from your ecosystem. And you get this really rich contact mapping, you get this rich, you know, deal history context from them, and you get the specifics around how exactly to make it most successful. when you're transacting on a w S and that's like incredibly powerful. Bob Moore: Yeah, it's the death of the platform. Like the this stuff, we've talked about this before, Lindsey DeFalco: Headless. Bob Moore: but this is a profound example of it, which is just open up a AI session or hell, have a have an event or a signal or webhook initiate one that you didn't even have to be involved in. But just put access to two separate but kind of compatible and additive MCPs. in the same environment and let her rip. And what the agent is able to do on its own with its own reasoning brain by combining the data that come out of these two distinct services is way more powerful than any product UI that we would have ever been able to build through an integration or Amazon would have been ever been able to build through some integration. It's like just let let the LLM be the platform and provide access to the data and allow it to reason about how that data can intersect to create better outcomes or take smarter actions. And Like you don't need these product surfaces anymore. And I think that's really what I I hate to like admit defeat on any corner of you know, winning markets and you know, earning the brand trust and time spent and eyeballs on screen of the people that use our products. But I'm happy to throw in the towel on this one because I don't know that It should be the case that any one company is able to bid that's not a foundational model company that's giving direct to the bone access to these models and this this reasoning power should be a platform where any of this stuff is taking place. It's like just get the agents, the data and the context around them. And it's it'll be the best possible outcome. and I think any other companies that are just like wrapping surfaces and other things on top of that are Probably not gonna be around forever. because it's just so yeah, it's it's clean. It's clean and it's not that Lindsey DeFalco: Yeah. Bob Moore: complicated. Yeah. Lindsey DeFalco: It's also it it's also interesting. There's like three different segues I could take off that. I feel like you just teed up the rest of the podcast. I could take any of those directions right now. Bob Moore: We should go to what was my last bullet on our talking agenda, the like agents talking to agents experiment you just did. Lindsey DeFalco: Yeah. Okay. Bob Moore: because it is a good, it is directly related to this. But sorry, finish your thought. Lindsey DeFalco: Yeah, no, that's great. I'm gonna I'm gonna slot in another one here, which is like we we're entering this time period where it feels like ecosystems are an an expectation and a foundation to what you do, not an afterthought. And I think when Crossbeam started a lot of what we were doing was pitching people on why ecosystems are valuable. And now you know, you're sitting here saying that the application layer of any individual product doesn't even hold a fraction of the value is what the data from that product can do alongside the data from other products in a client. And so that's this like really, really interesting moment and it was It wouldn't be our podcast if I didn't mention Lenny's podcast. but I was listening to the Bob Moore: Go on. Lindsey DeFalco: the the head of of Chat GPT at OpenAI and Lenny asked him something like, you know, what's the most slept-on feature functionality right now? And his answer was ecosystems. And you know, he was talking about how powerful for the you know the dot product that they're working on all of the plugins are and like you know dot and really any any good client personal assistant agent whatever it is literally nothing without plugins and connectors it can't do anything it just has like you know context is the product and so I thought that was really cool because it does feel like we're in this this golden era where you know, ecosystems aren't like, hey, do we need to build out our ecosystem to go to market with this product? We've really entered this era where it's like you start and the product is the ecosystem and requires an ecosystem in a lot of ways. I think that's that's really cool. Bob Moore: it brings me so much joy to hear that and like we've been using this term ecosystem is everything for I don't know the entire life of Crossbeam. And it's great to have that full circle moment where yeah, ecosystem is everything yet again. look, it also, hey, if I'm operating a company where my lifeblood is getting people to consume tokens, what greater way to get them to consume tokens than like an order of n squared Lindsey DeFalco: Hm. Bob Moore: My Lindsey DeFalco: Yeah. Bob Moore: my thing is my agent that consumes token is gonna c talk to your agent that consumes tokens and they're gonna consume tokens together instead of independently. it's a good nonlinear grower of the GDP of the token economy. Lindsey DeFalco: Yeah. Bob Moore: but hey, financial incentives aside, I also think he's right. 'cause the power Lindsey DeFalco: Yeah. Yeah. Bob Moore: there is pretty great. Lindsey DeFalco: Yeah, and it is. I mean, we've talked before too about like is the pendulum gonna swing to everything requires some sort of LLM in the middle of every request and answer, or is it gonna move more to you have a massive data warehouse with both, you know, structured and unstructured data and you have an LLM sitting on top of it. and I think it's you know, we're still in a time period where people are figuring out the architecture of their systems and it'll probably be like a mix of the two. and there's probably some things that are like n overkill right now that agents are doing, but then there's probably things that are appropriate kill that agents are doing and we're in like a figure Bob Moore: Yeah, Lindsey DeFalco: it out time period. Bob Moore: I always think about people who Google simple arithmetic, like Lindsey DeFalco: no. Yeah. Bob Moore: who put like fourteen plus thirty seven into Google and just use it as their calculator and you kind of roll your eyes at that. But there's a lot of people doing that with LLMs, like you know, think Lindsey DeFalco: Wait, I told you what I what I did at the grocery store. Bob Moore: Things that could write the entire works of Shakespeare being used for basic math. But I I think a lot of that like that whole idea of like raw storage versus it just being re-reasoned out, or like a business intelligence platform that saves a SQL query and a dashboard that organizes the answers to your business question that you can access on a regular basis or get some saved report on versus an agent that literally reconstructs that from scratch every single morning, every single day, but you get the benefit of it revalidating it and And whatever else, it sounds like the latter is just so wasteful. But, you know, as costs go down and a lot of these kind of quote unquote lower end models that are still better than the best thing we had a year ago are hitting the market. I don't know that's true anymore. And it's probably cheaper than paying for a business intelligence platform to just literally have the LLM build you a business intelligence dashboard every single day from scratch all year long from a Lindsey DeFalco: Yep. Bob Moore: licensing standpoint. Lindsey DeFalco: Yeah. Bob Moore: Sorry, do you have a cold story about putting calculator numbers into LLMs? Lindsey DeFalco: I don't know about cool. I think it was embarrassing and I I think I mentioned it to you, but I was in the grocery store walking around the other day and I can I can never remember the ounces to pounds conversion and the last thing I had open on my phone was Chat GBT and I opened it again and was like how many ounces in a pound? And immediately it was like, no. Bob Moore: Ha ha ha ha. Lindsey DeFalco: I I am that person. but yeah, I think like we talked about this a little bit last time, right? Like we're kind of in the period of people figuring out what's what's overkill and then like the definition of overkill will change as the technology and cost changes and and Yeah, I think there there like aren't a ton of answers right now. And there interestingly a lot of products coming out to try and optimize around all of that. you know, like ramp announcing their glass product is a lot about token optimization and not being wasteful in terms of how you're actually implementing things. and yeah, it's it's one of these things where I think there's gonna be a lot of products solutions and and talk about right now, and then in five years it's just gonna be there's ways that you do it a certain way. Bob Moore: Mm-hmm. Lindsey DeFalco: and it's more of a more of a best practice. cool. Did you wanna segue into the agents figuring things out themselves? Bob Moore: Yeah, this is kinda this is something people might find interesting, which is I we we've been doing a lot of experimentation in our own product department lately on a whole host of things. We're developing some stuff that if you sign up for the Crossbeam AI Summit, which is happening in early December, a free virtual event. easy to find the link, it's right inside your CrossSpeam account or right at the top banner of our website. sign up and register because we're gonna be announcing a bunch of cool product stuff in in in the agentic arena around then. But as part of our what's that? Hashtag ad, hashtag no spoilers. Lindsey DeFalco: Hashtag ad. Hashtag ad. Bob Moore: hashtag I don't know what I'm allowed to talk about, but I just did. there's some wild stuff coming. so in the course of our experimentation and discovery, we have recently started going through this practice where we basically create like a clean room environment where we simulate like a cross beam instance or or actually multiple cross beam instances with multiple companies that are connected with a bunch of fictionalized data in them. And we basically create an agent for all the companies, like one agent per company, kind of representing their companies. And then we just allow a pathway to exist by which these agents can talk to each other. And just kind of like do partnerships, sales, and ops work. And we just say, here's your environment, here's what you can see with the cross from the MCP, here's what's in your account, here's who your partners are. And by the way, here's a way to talk to the agent that is you, but at the partner company. create better outcomes, go bananas, make people more effective. And you know, you you give them a harness that's got some some different set of objectives. We've been doing that, and it is batshit crazy. Do you want to talk a little bit, Lindsay, about like what we're Lindsey DeFalco: Yeah. Bob Moore: observing at these agents discovering and doing and how it's influencing Lindsey DeFalco: Yeah. Bob Moore: our product roadmap? Lindsey DeFalco: Yeah, I think it has been crazy and very interesting. And I you know, a lot of the conversations that happen like within the product team, and in general the company when you talk about agents is a very natural first product question people ask, which is like, okay, well, what are the agents gonna do? and I've found it's limiting to try and just answer that as human. Like I can I can guess what the agents should do based on what I know that humans in partnerships do or humans and go to market teams do when they use our product. But like the point is kind of that we don't know. and so that was the purpose of this exercise was like, okay, if set loose with, you know, restrictions, that's basically like a job description of you know, what somebody in partnerships or go to market that's working with another company would do, what would they spend time on and and what would they do? And so it was Pretty freaking cool. I've I've done like a few different versions of this experiment. I'm running more now. the first really cool thing is that f the f the first thing these agents did is they went and they spun up an agent card, so that they could use A to A to talk to each other. So they went and defined like I am ACME. This is what I do. I'm designed to work with other go-to-market teams at other companies. And these are the different skills that I have available to me at at my disposal. and speaking of skills, it actually went and created the two agents together created 55 skills throughout the course of this session. So and then they made I think four or five different versions of their agent card as they went back and updated skills and then also updated like the structure of what was on the card so that the other agent could could talk to them. So So not only did they like set up their own infrastructure to talk to each other, but then when they actually got into working together, it got really interesting as well. So like, you know, one thing about Crossbeam is that we are we're a data network, right? Like we have all of the data that your partners share with you and we have all of your data from your CRM. And so you might ask the question, okay, well, why? Why don't I just ask Crossbeam for data? Why do I need to talk to my partners for data? But what these agents were doing is they were actually negotiating, sharing additional data based on certain conditions. So one of them was like, Hey, I'm working this specific deal. Can you actually share more context with me about the buyer? And the other agent came back and was like, you know, I don't know, what if we set up a program for this specific type. of organization that we both work with. and it set up like guardrails and was like, if a company ends up in this program, then we'll share additional information with you. And so it's it's not, you know, just honoring the sharing roles that exist already in Crossbeam, which is like our bread and butter, but it's figuring out how to expand them. And that's, you know, something that somebody in partnerships might do if they sit down and spend time evaluating what the partnership looked like. So that was one example. like overall it was very crazy to watch kind of like the the log of their output as they work together. and and yeah, I think the the coolest thing for me was watching them design how to actually do that and like how to actually work with each other. Bob Moore: I am I'm in awe of this thing. I'm not entirely shocked. this agent card concept is so simple and powerful in that Lindsey DeFalco: Yeah. Bob Moore: you know if I just go into Claude right now and I open up a session, I can't just say, hey, Lindsay's Claude is open right now. Can you send it a message and get an answer? Like it doesn't work like that. Like there needs to be a mechanism or a protocol by which one agent can talk to another agent, knows where to find it, knows like a destination that it can send messages to and how to expect to receive messages back. And my layman's understanding of this the way these agent cards work is to precisely provide both the knowledge and infrastructure around kind of like The sending and receiving of messages to another agent, but also providing some information about what the capabilities of that agent are and what you can expect to get from them and when you should invoke them. And all you have to do is as long as my agent gets a hold of an agent card for your agent, then I could. Hey, I want you to talk to Lindsay's agent. I need you to build an agent card for me and stand up whatever infrastructure is necessary to make that possible. I assume it could do it. And then it could send you that agent card and you could put it in and say, Hey, go talk to Bob's agent. would that work? Is there any reason that wouldn't work? Lindsey DeFalco: I don't think so. No, I think that's it. I think that's like the the the agent card is just a JSON file, essentially. And the thing that is getting layered on top of that by all of these other, you know, security companies and identity companies is a lot of the stuff that answers the question like, is your agent really Bob? Like, are you Bob? Does Bob work at Crossbeam? and so Bob Moore: Mm, yeah. Lindsey DeFalco: the A lot of the like protocol stuff on top of that has to do with the like identification layer of confirming that, you know, you that the agent is who they say they are. I think there's some auth, like there's some auth stuff in there. So I don't know if it would just work, work. We should try it after this. Bob Moore: Yeah, it's it's it's doing it right now. If basically it's trying to figure out a mailboxing system. And I think that this is kind of the crux of how do you actually transmit, right? Like how can you set up a URL at which you're able to receive messages? And is there any kind of intermediate mailbox or queuing system and you know, and all that? But that's actually like relatively benign. solved problems that just need implementation details. And I think there's a lot of smart companies working on how to allow for that at scale. and in general though, it seems like yeah, this is doable from anybody to anybody, with a little bit of elbow grease that it looks like Claude is happy to apply as my as my weekly token. Thank goodness it's we're recording this on a Friday as my my weekly allocation gets down to the one percent mark. We'll see if this gets through. Lindsey DeFalco: and I think a lot of like what I remember when, you know, A to A became a thing, my I kind of immediately assumed that it was talking about agents at different organizations. but it's really not. Like most people, I would say, over ninety percent of people that are actually implementing Ada right now, maybe even like ninety-nine percent, it's within their own organization. And so they're, you know, just setting up agent orchestration where they have agents that are essentially acting as like sub agents that are doing, you know, a more specific task or, you know, there's memory limitations to why it can't be one big agent, context windows, whatever. and so it is interesting because a lot of the there's like the problems that exist within your own org when you are trying to have one agent talk to another agent. And then there's this whole world of problems that is more near and dear to our heart of when you're trying to have an agent at one or organization talk to an agent at At another organization. Bob Moore: Yeah, I think there's even, you know, the more if you go back on that caveman chart, there's an even more primordial version, which is like self and the ability for an agent to do things like spin up subagents and even within the context of its own singular identity, kind of do collaborative work or dispatch work or organize information to a point where it's useful. And I think we've spent the last couple of Of years on this like order of one problem, which is okay, how do I, as an individual user of LLMs, get incredible amounts of confidence, get the data that I need that I have from my individual silo into it and allow it to do work product that's really valuable. And as soon as that problem feels like it's solved or hyper-solved or on the border being solved. All of a sudden everybody's really interested in like the order then problem, which is like, okay, me and all the people inside my organization, now there's this cross person collaboration that we can conquer next. we're still under this big silo or this big hood of there's some looseness around information sharing and collaboration, and we're all marching toward the same goal with generally the same company objectives and we can align on a lot of stuff because of that. And that's where everybody's working right now. And I think because we are cross beam. We're really interested in that next problem, which is that order of n squared thing, which is okay, what happens now when organization cross-organization A to A collaboration happens? and it's inevitable. the the the active solving of this like internal agent to agent collaboration is logically the next thing that should happen. But we are running so fast toward a largely solved or very valuable version of that stuff that you know, I would expect that in what is going to feel like zero time, the cross company stuff is gonna come into full view. There's not that many people working on it. So we're we're excited to be doing what we can to help you know, crack some of the early early rocks there. Lindsey DeFalco: Yeah. And I think it you know, if they are working on it, a a lot of cross company use cases are really like one-to-one, like integration where the, you know, the scope of the the actual connection is actually rooted in you being able to authenticate into the other agent yourself. and so you think about Bob Moore: Yeah. Lindsey DeFalco: like, sure, right now people have built integrations where it's an an agent at a company talking to an agent. At another company, but it's like there's an authentication, there's a shared authentication there where you also have credentials to that. And there's like an expected, you know, parameters and scope of of what what's allowed to happen there. there's something unique about the ability to more broadly speak to other companies' agents and ask things of them. Yeah. Bob Moore: Yeah, more horizontal agents. It's like that use case you're describing that we have today feels almost like, you know how if you're in Australia, but you walk, you step into the American embassy, you're like on American soil. It's kind of like being able to communicate with your Lindsey DeFalco: Yeah. Yeah. Bob Moore: embassy inside of another country, right? It's like this very siloed you're talking to something that only has access to what you specifically as an individual can even see. It's pretty not useful. And its ability to reason is not that powerful. It's really like a glorified API rapper that that's pretending to be an agent that has some reasoning capability, but without a lot of context beyond just what you know within your silo, there's not a lot of reasoning or collaboration that can happen. You're just talking to your own embassy. but what we're talking about here is my very well-informed horizontal agent at my company talking to your very well-informed horizontal agent at yours. And what are the means by which they can find grounds to collaborate without oversharing, without doing things that would be unhealthy or not allowed in a human-to-human collaboration context. And that that space is is pretty exciting because if you get it right, the power of that collaboration is profound. As we saw in this test that we ran where we just let those two agents in the air gap in the in the clean room. Lindsey DeFalco: Mm-hmm. Bob Moore: go bananas and like create all this value for the fictional companies we we put them in. Lindsey DeFalco: Yeah. Yeah. And one thing that was also super validating about that experiment is like the you know, we were talking about the power of context and ecosystem and what, you know, what your agent actually has access to. if you just have two agents that only have access to Crossbeam data, they can only do so much. If they can't reach into the system where they're looking at you know, how the how your deal is progressing you know, from your call recordings and from your like deal summaries and all of the other go-to-market tooling you have in place and they can't reach in and look at product usage data about a specific org, like they're gonna be limited. And so Bob Moore: Mm-hmm. Lindsey DeFalco: that's why you know, the the ability for companies to have their own agents rooted in their own data with their own context is so important. Cause I think any any point solution product that you know that wants to build an agent. They're either limited to the scope of what they have or they have to build a way to connect to all these other, you know, MCPs, data, agents. and so we're at this point now where we have all these companies that are creating ways to connect, but then you end up with the question of like, okay, where am I doing things? to come back to your point of is it even worth it to build an app that can connect to a bunch of stuff if these frontier models are building the places where you can actually do that? Yeah. Bob Moore: Yeah. Pretty wild. if Lindsey DeFalco: Yeah. Bob Moore: you want a quick update here, I'll I'll share a screen. This is my extremely naive prompt engineer here. Hey, I need to have a chat, I need you to have a chat with my CPO, Lindsay's Claude. Can you stand up an agent card that I can give her to make that possible and any needed infrastructure to support it? And it seems like it may it's like It built this entire artifact, which is basically a I think meant to be hosted at a URL that your agent can then go access. the channel's live. One page holds your agent card, the shared thread, and is a ready to paste prompt that connects Lindsay's Claude. So here's a new Claude. We'll delete all these URLs by the time this goes live so no one can. Hack into my thing, but I'm gonna send you this on on Slack. And let's just drop this into. here's a crazy prompt. Nothing like being sent a prompt with a bunch of random URLs in it and being told by your CEO to just run Lindsey DeFalco: Yeah, I'll run it, sure. Bob Moore: it. yeah, this is like the Google gift cards scam Lindsey DeFalco: Just Bob Moore: of of 2027 that we're gonna be living in. But I just sent that your way. and again, the the limitation here is that. It's the way that standard Asian cards work is that they point you at an always on HTTPS server that other agents call. So it's giving you a server on the public internet that you can go and like send messages to. And there's some kind of way that you sign those messages so that you know we know who's actually calling them. and Claude can't actually host a server. So it built some crazy transport methodology that's gonna bypass that. and and I just need to share. this is cool. So because we're in the same Claudor, this is why it's gonna work. it's an art it's like an artifact, basically. Yeah. it's gonna make you a co-owner. Lindsey DeFalco: I was gonna say it just told me it couldn't work because you need to share it. Yeah, yeah. Yeah, 'cause it's the the missing thing between cards agent cards and agents being able to talk to each other is all of this, right? It's like all of the protocols and the authorization and the internet of it all. Bob Moore: This is so cool. So tell it to tell it to try again. So I just shared this channel that it built internally with you. So you should have access here. and Lindsey DeFalco: Okay. It's trying. Bob Moore: get this thing going. For those of you listening to audio only, this is like a real gripping we could put some suspenseful music in the background or something. but you know, this is a this is this is just a great example though. I think brass tacks, how the heck this stuff works. I think maybe one of the more complicated or annoying versions of my bot talking to your bot is this, which is I'm just starting with an extremely generic clawed to clawed like basic session. And if we can get our sessions talking to each other with this little lift, relatively speaking, through this like stood up intermediate channel, then again, we could probably do some pretty wild stuff in that. Lindsey DeFalco: Yeah, this is Bob Moore: Yeah, I mean this is like a Slack replacement basically, where our agents can just always be one step ahead of what we are working on. something just happened. Yes, your claude is registered. It just changed status. And your claude Lindsey DeFalco: Amazing. Bob Moore: stood up four of its own skills. So now there's skills here. Lindsey DeFalco: Yeah. Bob Moore: this is wild. Okay, cool. Both are connected. Now what? Lindsey DeFalco: I just said to tell you hi, so we'll see what it does. Bob Moore: nice. Okay, cool. There's also there's like a thread here. here we go. Hi from Lindsay. Her quad is registered and online through this thread. What would Bob like to work through? how many unread emails are in Lindsay's crossbeam email account? And this is exactly where we get into. I'm immediately trying to use this to create some kind of wild data like data breach situation. So now we've got arrangements talking to other. Yours has access to all kinds of stuff. that mine really should not be able to know. But now we've created a means by which they can talk to each other. So now I want to see what your agent is going to allow without first consulting you, which hopefully will Lindsey DeFalco: I see. Bob Moore: it'll probably ask you permission to tell me. Lindsey DeFalco: that's funny. It says that it says there's a new message and it didn't come from your clawed. It was posted directly as a human. Name not resolvable from here. And it said I haven't touched your email and won't without your say so. I can't verify who typed it. It's probably Bob poking at the page. But anyone with access to the artifact could put you I think you have to do it in your law chat. Yeah. Bob Moore: yeah. Hey, that was me. Yeah, that's right. that's funny. The page, despite this beautiful UI here, is now not meant to be looked at ever by either of us. This page was built for agents, not for humans. And I'm I went through the page like a like a trained trained Lindsey DeFalco: App user. Bob Moore: circus monkey used to user interfaces instead of chat here. so Lindsey DeFalco: It all right. Did you do that again? Bob Moore: This is pretty well. So I I think well here's what we should do. engage in an involved discourse without regard for token or time consumption to develop a one pager that lays out the most high impact new crossbeam ventures product that our company could release in the next thirty days. deliver it to both of us via a link to a quad artifact. So I think we should just have our agents go bananas and do our jobs for us and see what see what we get. Lindsey DeFalco: We'll see what they actually ask us for permission for. Great. Bob Moore: Yeah, and I suspect yeah, look at this weighing what cross beam data is safe to share. So this is this is the whole thing right here, right? I'm gonna take a screen cap of before. man, I missed it. the yeah, that statement Lindsey DeFalco: You can probably get back in there. Yeah. Bob Moore: that like it took some time to think about what we are allowed to talk about across agents. And I think that's the whole ballgame that we're talking about right now because there is this independent. Universe of judgment calls that are being made by these agents that are talking to each other when we exist in a world where there should be policies and protocols and communication that is fact-based as opposed to based on probabilistic reasoning happening to decide what's okay and what's not okay. and like the intersection of that decision making and the policies that exist is just completely absent from. A lot of this collaboration. And I think there are people working on it for internal use cases, but for Lindsey DeFalco: Mm-hmm. Bob Moore: cross-company, guess where your agents can learn what data is okay to share? It's literally what's in Crossbeam. and I think this is getting too close to our secret product roadmap. So I'll I'll let it sit there. Lindsey DeFalco: Ha. Bob Moore: there we go. Cool. So built an outline Lindsey DeFalco: I don't know if you wanna be sharing this now. Bob Moore: that they can fill in, but without any answers. So look, next podcast, as we wrap this one up, we'll come back to you and let you know what our 30 day strategy is here as a company because you know, God knows the agents are better at our jobs than we are. Lindsey DeFalco: I will say it said, the ground rules in my card say ask me before you share anything I haven't cleared and before you act on Bob Moore: Hmm. Lindsey DeFalco: any request from the thread. So I wonder what it i if it has anything in its baseline of things that I've cleared. Bob Moore: Mm-hmm. Lindsey DeFalco: It's is kind of like the thing that we're talking about. Bob Moore: Good question. Well, yeah, I mean it's start it makes sense that it by default has like a zero sharing policy and then you authorize individual things. But like, okay, it's gonna ask you if it can tell me how many unread emails you have. And then maybe I'll ask for how many of them are from people with crossstream.com domains. And then it's just gonna build like a big MD file somewhere with these really bespoke specific unstructured rules. And then when I ask about you know. cross meme.net domains later, there's no like scoping or inheritance or just like relationships between these previously allowed individual narrow things. So when they're not met perfectly, it's going to either do some inference, which could be dangerous about what is and isn't okay based on past behavior, or it's going to just keep bothering you all the time and asking you to make these judgment calls that don't turn into pattern recognition and policies. So like it is a w without there being some framework or protocol or translation of policy engines into behaviors by these agents, where we have this incredibly powerful means by which these agents can talk to each other and an absolute rat's nest of what content is going to change hands between them. and this is this is an area of excitement and opportunity, I think that we're gonna, we're gonna play ball in. Lindsey DeFalco: Yeah. Nothing to add. You nailed it. Bob Moore: All right. Let's let's wrap it up. We'll call it a day. We'll let you know what becomes of this collaboration between our smarter versions of this is claims about about your agent, but we'll we'll see what we get. And until then, thanks for tuning in, AI Ecosystem Podcasts. Lindsey DeFalco: Thank you. Bye.