Latest / The AI Ecosystem Podcast / Agent Fatigue: Spotting “Fake Agents” and the arrival of XAA
Transcript
- Lindsey DeFalco: Hello. Bob Moore: Hey, welcome everybody to another fantastic edition of the AI Ecosystem Podcast. Bob Moore here with Lindsey DeFalco. How's it going, Lindsey? Lindsey DeFalco: Great. No complaints. I'm very excited about our topic today. So Bob Moore: We've got some good some goodies. it's been a couple weeks since we laid one of these down and as always the world has shifted drastically. There's new vocabulary, there's new three letter acronyms, there's new trends and things we're we're dealing with and the the name of the game today, Agent Fatigue. Did we make that up by the way? I think you coined that. I don't know if that that's like a whole thing. All right. Lindsey DeFalco: Yeah. No. De definitely not. Bob Moore: Well, I guess like the when I think about it, I think of it as like that may not mean what you think it means. how do you define it? Lindsey DeFalco: Yeah. I was gonna say I think there's like A few different aspects here and like different angles we can look at it from. So one is everyone's calling everything an agent. so I have this bit I do on calls where every time I say agent I put it in quotes and I say I put it in air quotes because everyone defines agent differently. so I think part of it is that every single company is launching an agent. And are they actually agents? Who is the authoritative definition on what an agent is? and that the breadth of what that means is pretty wide. I think that's Bob Moore: Yeah. Lindsey DeFalco: That's one angle. I think another one is also well Two more. So another one is also like tokens. We're I think we're out of the token maxing phase and we're deeply into the token optimization phase at this point. and what do agents mean for that? Do you actually need an agent for this? Is it actually just a workflow or an automation? Are those bad words now if you can have an agent? and that whole theme of of work. And then I think third would be the governance and orchestration of agents and it kind of combines the first two in that if everyone has an agent and everyone's calling everything an agent, you need to be able to govern who can access these agents and how they all work together. Are they sub agents? Are they super agents? And so the whole architecture piece of this in terms of trying to put together full solutions at, you know, say an enterprise company gets gets really complex and you have this this agent fatigue I think where you've got all these different solutions that are all wrapped in an agent and now they all have to work together with appropriate governance. That's how I think about it, is those three buckets. Do you have another bucket? Bob Moore: no, I think it's I think it's fair. though those are the I I think in particular the everybody's calling everything an Asian is an interesting thing to drill into. I I do think there's two dimensions to this that line up with yours. Like one of them is just slapping the word agent on anything that feels like an automation that may or may not actually even be powered by an LLM at the core. and the deconstruction of different classes of of solution that fall into that agent bucket is pretty interesting and we should do that. I Lindsey DeFalco: Mm-hmm. Bob Moore: think the other thing is among all the companies that actually build what I would say is the true purist definition of an agent, which is something that is ultimately kind of LLM powered. It is using LLMs to Process the overwhelming majority of the input that it receives. It's creating novel outputs and is operating autonomously, not just being triggered by human input as like an initiator, like that real class of agents. There's an agent fatigue there too, because outside of the foundational model companies like ChatGPT and Claude's like actually being the backbone of agents, there's a whole universe of companies that are actually trying to. play ball in the agent game and have you build agents on their platform. so like this is what Salesforce is trying to do with I'm gonna mix up what's agent exchange and what's agent force, but I think Agent Force, right? You b you build agents that actually live and host and like the LLM calls are coming from Salesforce. and certainly there's you know kind of middleware-ish They probably wouldn't want us calling them that, but middleware-ish technology platforms that sit in between the foundational models and the companies that use the agents internally, like Glean, for example, that are kind of like agent engines, basically, and knowledge aggregators. But ultimately they become kind of like the way you can run an agent. If you go into Gong, I think you can like just run a Gong agent. And obviously the customer support platforms like power agents that are customer-facing. So There are actual real agents, not just things that are labeled agents, that exists somewhat prolifically across a universe of tech products. And it forces this question of do we all really actually want to have our agents distributed out across all these different tools for all these specialized tasks? Or is there some merit in being honest with ourselves and saying the important thing is actually that the agent has access to information and context that's relevant to what it needs to do, but it doesn't necessarily need to live inside of the tool that owns that context, right? And there's this like there's this coupling of who owns the data, whether it's your partner ecosystem data or your CRM data or your support knowledge base data or your internal company emails or whatever, who owns the data and then who owns the agent. And it seems fairly obvious that Those two are very distinct layers, that need not be integrated with one another and in fact are probably impaired by being integrated by one another with one another. And yet a lot of these companies, especially ones of large scale, don't want to give up the real estate of being the agent runner. So they're using their data leverage to try to become agent platforms as well and be the place where the agents are running. And they're not all gonna win. Lindsey DeFalco: Yeah. Bob Moore: In fact, most of them are are gonna lose. and it's an interesting question, like wha what's the natural place for these things to live? It it's it strikes me that either the foundational models or some dedicated layer designed specifically for building incredibly powerful, robust agents with very diverse context available to them, probably like the gleans of the world, make the most sense. but Lindsey DeFalco: Mm. Yeah. Bob Moore: Everybody's out there fighting that fight right now. Lindsey DeFalco: Yeah. And I think it's like I mean, the question is like where's the data and also where's the governance? And I think that's like a lot of what the conversation is right now is like, okay, you want to build your agents where you also have control over who can access them and when and how. And so that's like the thing about building out, you know, scattered agents and different systems is then you're beholden to the access in that system and it's harder to manage in one central location. And so that's also, you know, an interesting thing with like the the agent forces and and the gleans is there's like some level of, you know, permissions and and access baked in Bob Moore: Mm. Lindsey DeFalco: there. And yeah, I do. I think it's like we're in this weird time period where no one is I won't say no one. It's like no one and everyone are claiming like you need to do all your work here. Everyone's like, hey, you can do your work here or you don't have to do your work here. You can do it somewhere else. You can build an agent here or you can build an agent somewhere else. But it's like Bob Moore: Mm-hmm. Lindsey DeFalco: how much you know context is actually available to you in the place where you're building your agent and then can you actually control at scale you know, who can access it and how. Bob Moore: Yeah. Yeah, it's a good one. I mean, like, look, if if if I'm any one of these companies, why would I not just create a thin wrapper to one of the the foundational models and suddenly have an agent? and hey, look, I've got some of the most valuable data that your agents might need, your generalist agents might need, and I can build in hard wiring in the background so that you don't have to worry about Flaky MCP connections and authorizations and things like that. And like you know that that data is always always going to be there. Like if I'm gong, your gong data is always going to be there with these agents. So if I'm Salesforce, your Salesforce data is always going to be there. And then you kind of build out around that. And if you believe that you are the center of your ecosystem, then it makes logical sense that you would actually be the host of the agent. the the technical overhead of Building at least the simplest versions of these is quite low. The hard part starts, yeah, coming in when you get in these enterprise deployments and you have permissions and all kinds of other stuff, which again, if you've got them baked into these big foundational platforms that are like, you know, systems of record, then hey, layering agents on top of that maybe can inherit some of that rights and rules and permissions to make that part of it a little easier too, right? So like I I guess this is a place where. incumbents potentially have the ability to benefit or at least make an argument in the enterprise where there's like really these really complex complex needs. But Lindsey DeFalco: Mm-hmm. Bob Moore: you know it gets us right back where we started, which is a bunch of tools that probably do a lot and are increasingly horizontal, kind of also moving vertically up the stack now. to to be infrastructure, like agentic infrastructure. Lindsey DeFalco: Yep. yeah, I'm I'm sent multiple new companies a week that are like agents as infrastructure, infrastructure for agents. We help you build Bob Moore: Mm-hmm. Lindsey DeFalco: agents in twenty minutes, not two hours, or not twenty hours, like it the yeah, it's definitely a becoming a crowded market already. I think there's in the governance area there's an interesting speaking of new acronyms, Actually it's not that new, but speaking of acronyms, so like Okta, for example, has this feature they call cross-app authentication. XAA for shorthand. And it's based on a standard protocol, so it's not you know it's not not specific to like any any other IDP can also support this. But what it means is that the Bob Moore: Thank you. Lindsey DeFalco: industry has centralized around a way to actually manage MC. Connections from a central location. So this is like incredibly compelling to me for from the Enterprise story. and the best way I've found to describe it is like imagine you're trying to roll out a new connector to your whole team and they all use Claude. So right now you have to tell them, okay, every single person go authenticate this connector, and then it will get access, and then you can use it. And so what this does is you log into Claude, and the fact that you have access to To the connector and what access and it actually being authenticated are already all handled by your IDP, like an Okta. And so you log in and you have your whole set of MCPs and tools. And so what that gives people is this like one central place to actually manage access to different applications in the MCP era, which is, I mean, I think that's just gonna be like a table stakes for for any of these enterprise level deployments. Bob Moore: this is super cool. I I made the joke to you the other day. I was like, you know, I feel like one of those grandpas that's gonna be like, kids when I was your age, we used to have to log into our own tools in our MCP environment. And it is pretty wild that right now, like, every single person at our company has to log into NADN, has to log into Snowflake, has to log into Salesforce, has to log into Crossbeam. all just to like individually connect their individual accounts to the MCP servers of those services in order to have their their claude instance be able to reference all those things and connect. It's so obvious that yeah in the same way that SSO exists, like you should just be able to at an organizational level, like bless the MCP for any of your employees that have an account and just have those things line up and align. and it's cool. It sounds like Octa and Anthropic have been working like really closely on on co designing this and it's very likely to become the standard. and it should exist and it'll save us a lot of headache, frankly. And me as an end user a lot of headache because I have like log I have to re authenticate linear like fifteen times a day. No shade Lindsey DeFalco: Ha ha Bob Moore: linear, we love you. but it you know, it's an annoyance and it's unnecessary. So these things are getting getting pretty pretty cool. And we will obviously lean into to participating in them. The other cool thing is that this like it just reiterates and builds more infrastructure behind the fact that like people were so concerned about user seats dying and going away, and that any Lindsey DeFalco: Mm-hmm. Bob Moore: seat based licensing was was a problem and Interestingly, this is something that like brings seats like right back to the to the forefront and kind of you know making sense as an atomic unit of value. because if you're as long as you have human employees and they're leaning on these agents, they they are seats and they need authentication and the tracking against this is going to be very linked to an individual in a way that's actually it's gonna make it easier for companies to you know make sure that they're collecting on every single human that's benefiting from their product, and not hide behind multiple people logging into one thing or something like that, right? Like when you're backed up by SSO providers, it it's pretty hard to do one to many. it's a big Lindsey DeFalco: Mm. Mm-hmm. Yeah, and I Bob Moore: thing. Lindsey DeFalco: think the th the thing with the theory about seats going away is that that theory severely underestimates the amount of like app specific permissions that exist. so even if, you know, you're just getting data out of that app, there's still, you know, controls that are baked into what data you can get. And you know, no admin wants to specifically manage that across hundreds, thousands of apps in one place. They don't know what that app does. They just want to inherit those permissions. permissions and still have the ability to have them be per user because, you know, the authentication is technically still going through through the user. So yeah. I do think that's a funny we're we're watching all the the cycle of like fear tactic article. the most recent one being that AI is going to kill everyone, which I don't think we're gonna get into on this podcast. But we go from fear mo fear mongering article to watching it actually play out in real time and then like settling on the reality and the user seats one has been interesting 'cause I think we're like, we had all this seats are going away, everyone adjust your pricing. and now we're kinda settling on this like, no seats are still they're still a thing. They're still important. Just for Bob Moore: Yeah. Lindsey DeFalco: a a slightly different reason. Bob Moore: Yeah. yeah, it is, you know, since our last podcast, the the market, the SAS pocalypse affected market has also come rallying back quite a bit. and we've seen Lindsey DeFalco: Mm. Bob Moore: like, I don't mean to keep mentioning Salesforce 'cause there's a bunch of comps here, but Salesforce had a, you know, a day where it went up twenty two or twenty three percent in a single day and recovered a lot of the it's basically back to where it was in you know, in January, before we went through like the the doldrums of this like the SAS world's gonna end kind of kind of storytelling. It's only eight percent or so off from where it was at the beginning of the year. and yeah, I think a lot of this like look, they announced Claude Force and like a really deep partnership with Anthropic, but The market's not actually dumb enough for that to be the only input to that, right? I think the other thing they did is continue to report out on steady earnings growth and continue business performance. And like most of the stocks in the in the broader the broader market have kind of done similarly, right? Like Snowflake is another one. Snowflake's up fifty percent on the year. and you know, there were there were certainly questions if you look at it, you know, in the from late 2025 to this summer, it was down God from in the like the two hundred eighties to the one forties. Like it was it was down to half of it. It's lost half of its market cap value in the one year less than one year time period, and it's all back now. and nothing changed except the company just like didn't tank quickly. Lindsey DeFalco: Yep. Bob Moore: in fact it continued growing and A lot of these companies tightened their belts and, you know, improved earnings performance and did a bunch of other things. But I think it is it is a little bit of the like haves and haves nots around folks realizing that there's like there's still a win here. which is pretty pretty interesting. Lindsey DeFalco: Yeah. Well I'm sure we're about to hear a lot of exciting stuff from all of those companies during conference season that's coming up. we're we're snack in the middle of. There's Bob Moore: Yeah, I think it's very true. Lindsey DeFalco: gonna be a lot of big announcements about what what all of them are up to, I'm sure. Bob Moore: Definitely. well. Lindsey DeFalco: should we we started talking at the beginning and then got into the kind of governance side of things, but the whole Bob Moore: Well, governance is just such an attractive, interesting topic that everybody cares so much about that I think it just it lured us in before we could you know, mock Lindsey DeFalco: Just Spare, Bob Moore: people's fake agents that aren't really agents. yeah. Lindsey DeFalco: Yeah. Let's do that. Let's talk about agents, people building agents that aren't agents. I mean we could start with us. Bob Moore: Yeah, we build all kinds of agents that aren't agents. Lindsey DeFalco: Ha ha ha. Bob Moore: now we might o soon have agents that build those agents that aren't agents. but this is like the reason this is a hot topic for us is because we recently came to a conclusion that like we w we gotta tell like our our folks in you know, the the that hold the megaphone at our company, whether that's one to many or one to one, that we're not gonna go out there waving a flag saying, like, we have agents, turn on the cross beam agent. Because the reality is Not only do we not actually have agents, I don't think we we want to, at least not right now. for all the reasons stated earlier, we have this like really incredibly powerful and differentiated data layer that is an amazing input to the way AI systems work and are informed and and derive context about what to do at all of our customers. And they have their own agents. They're they're running. in these platforms that are either powered by you know Google or Anthropic or OpenAI or Microsoft or you know some of them use use folks like Glean who we've partnered with really closely. and those agents that are running on those platforms need cross stream data in order to do what they do incredibly well. And we're just not in the game of trying to compete with that. And I think it's a it's it's if it's a bad pitch to our customers to say, hey, I know that you have this very horizontal agent that has access to all your other information that lives inside of this other environment already. Why don't you also turn on an agent inside of Crossbeam that lives here too? And by the way, you're gonna have to connect all those other MCPs in this location and whatever else. Like, I just don't think people are gonna do it. I I think we s we serve a purpose as this very, very valuable data and, you know, I won't even call it infrastructure. it's ultimately information. that's that's very critical. And I think we want that information to propagate everywhere, but we don't need that everywhere to necessarily be stuff that we're just like white labeling LLMs over top of. so in us coming to that realization, it just has us have our antenna up to look more closely and realize that there's a bunch of people out there that also have phony agents but have have not steered away from that language. And I guess the common pla like Lindsey DeFalco: Yeah. Bob Moore: what is a phony agent? It's it's just like an automation, right? It's like Lindsey DeFalco: yeah, I think like Bob Moore: Yeah, an example would be like, hey, when somebody opens an email and it triggers, you know, an event in your marketing automation solution saying, Hey, this prospect opened an email, let the sales rep know that they opened the email. Or hey, maybe even draft an email and send it to them. Lindsey DeFalco: Mm. Bob Moore: and maybe the LLM is drafting that email, but If you've got a million people opening emails, you're probably just drafting the same thing over and over and over again. And it's a giant waste of tokens to even have the LLM inserted. So sooner or later, if you're smart, you just have a couple email templates that perform the best and just do some simple mail merge substitution. and yeah, it feels like what a lot of companies are realizing is these things that they've put LLMs in the workflow for end up at scale being. So ridiculously predictable in an enormous percentage of the cases that it's just a waste of money in tokens to have the LLMs do the work. Lindsey DeFalco: Mm-hmm. Bob Moore: and like, you know, whether it is things like drafting emails or routing information or like classify classifiers is a great category. like you have a data point coming in, a new person filled out a form or you know, you've got a new company that you're attempting to profile and assign to a sales rep or something like that, and you need to do all these classifications, these are all fairly discernible things you can do to figure out like, is it a public company? Is it you know, based in the United States, you know, what's its revenue level? Yeah, the LLM can tell you, but these are also facts, not in intuition requiring like Yeah, you can buy those data sets. totally, totally. Lindsey DeFalco: You could also buy the data set. And and and reference the data set. Yeah, it's I I think the the best litmus test for this is like find at your company if you think you might have an agent that's not an agent, it's an imposter agent, go find the most direct engineer at your company and ask them if this is an agent. and I have found that has been the most interesting conversations is as we're talking about building these things, if everyone who's actually gonna build it is like we could do this deterministically, why do we need AI? pretty good indication that that You don't need it. Bob Moore: Yeah, we frequent listeners out there will remember we did an episode somewhat recently on trace, which is the new product we have that is basically scraping the whole internet and identifying all these really interesting firmographic and technographic signals about companies. This is a perfect example where let's say you scrape every web page on the internet, you can absolutely send every one of those web pages into an LLM. And say, hey, look at the source code here. Let us know what technologies are clearly being used. Go check the jobs page and the job descriptions and parse out the text and figure out what tech they mentioned. And like the LLM will do an excellent, excellent job of it. but it will do it very expensively. and and Lindsey DeFalco: Mm. Bob Moore: you'll pay a fortune for each one. And the reality is we started doing that with Trace, just to grow a corpus of data and findings. And w by the time we had done a hundred sites. We were able to derive a bunch of these incredibly deterministic rules that that basically taught us, we don't even need to look for that with an LLM anymore, because it's so incredibly discernible. And there's this really interesting thing where good agents ultimately are almost like this cascading logic where if you're doing something at scale and you're doing it repetitively and it's the same family of tasks, but there might be unpredictable input sometimes. The right way to do it is in fact to say, hey, here is a set of deterministic rules that can be modified and continuously improved and iterated upon that determine the behavior, but there is always a fallback when the rules are not met that you can feed it to the LLM. And the LLM can either, you know, take a custom approach to resolving that input, disregard that input, or ideally do that. and modify the heuristic that is leading to all this deterministic stuff. And I think like because like yeah, instead of token maxing, it's like token minimizing. Customer support bots, great example. like the Lindsey DeFalco: Mm-hmm. Bob Moore: sophistication of these advanced models is just completely unnecessary in the overwhelming majority of cases. and in fact, such incredible simplicity as like simple kind of like case statements where you you route different outcomes to different pattern matching. Can actually do the job in 50% plus of the cases. And the amount of tokens that you save is wild. I just Lindsey DeFalco: Yep. Bob Moore: met a Y Combinator company yesterday that's doing stuff in this area. Whose name I think it's Nex. Is it Nex? Send it to you, Lindsey. yeah. Next, Lindsey DeFalco: That sounds right. Bob Moore: NEX, yeah. and their whole their whole pitch is like this exact thing, basically, right? It's like, hey, you've got agents handling very, very high volume tasks, insert us in the middle and we kind of build a harness around your agents that looks at all the inputs and the outputs and identifies extremely predictable things and basically like takes the LLM out of the loop and just builds deterministic structure. and it's brilliant. and everybody should be doing it. It's how trace works, right? It's literally how trace works with with processing data out of the out of the web for the firmographics right it's like LLMs generate a rules engine And anything that kinda doesn't get detected by the rules engine falls back to a queue that the LLMs then process to modify the rules engine. yeah, it's cool. Lindsey DeFalco: I I like that like that structure where it's you know, your main workflow is deterministic, but the LLM is like the backstop and is also continuously improving the deterministic portion Bob Moore: Yeah. Lindsey DeFalco: of it. And I think we've we've talked about that before with like matching, which like matching is core to, you know, Crossbeam's business. That's that's what we do. We we match accounts with each other. and the majority of what people ever want to do with matching is incredibly deterministic. and so taking our entire matching algorithm like dumping it in LLM makes no sense. It would cost an obscene amount of money for, you know, marginally different results. And so I also think we mentioned this. We had this really cool hackathon project done by somebody on our team named Donna who essentially did what Bob is describing works with Trace, which is like if there's you know, if our confidence is low in this match, then use an LLM to go reason and make a better decision about about whether or not it's a match, but kind of the same principle where like ninety, ninety five percent of it is is very well suited for something that's just deterministic. But the model can help Bob Moore: Yeah. Lindsey DeFalco: on the edges. Bob Moore: So then is that an agent? I in in some ways like I'm willing to give it a yes Lindsey DeFalco: Yeah. Bob Moore: as long as the ultimate backstop is the LLM. and you know, it's able to be kind of initiated autonomously by some non human initiator, like some signal coming in the coming in the door. Lindsey DeFalco: Yeah, I would say that specific thing would be a matching agent. Yes. That's like iterating on and improving the actual matching algorithm based on the what the LLM reasons. Bob Moore: Yeah, yeah. Lindsey DeFalco: I think that that's actually okay, so like what is an agent? We we talked about I think I talked about this this Lenny's article, shout out Lenny's, that walks through the three different stages, three different types of agents. And the first one is very heavily like deterministic automation. And the idea is that within one of the steps you're using an L L to do something meaningful. And so that's why this whole discussion that we're having right now and that we had, you know, for our own agents is interesting because if you have ten people and you ask them if what we built is an agent, you know, three might answer one way, seven might answer the other. and Bob Moore: Mm. Lindsey DeFalco: I had this very hilarious conversation with Claude specifically about this. And it was basically like, well you can do whatever you want. Like other people in your space are probably calling this stuff agents. depends on if you care, and where you're going with it and how much it matters to you that you're saying the word agent and riding the agent marketing wave. And so it is this funny like it on the one side we have all these actual like protocols getting built around capital A agents that are gonna talk to other capital A agents with like A to A. And then we have like all of these little lowercase agents. Ooh, maybe that's what we call them, lowercase agents. Bob Moore: Yeah. Lindsey DeFalco: that are you know, some of them more agentic than others. Does that matter? Kind of depends on your business and your market and your customers. and so it's this like weird space we're in it feels kind of similar to the co-pilot timeframe. Do you remember that one? Everyone Bob Moore: Mm-hmm. Lindsey DeFalco: had a copilot we still have a copilot so Bob Moore: Yeah, it's a really it's it's a great point. That was early early in the early after the Chat GPT moment where people just like dropped a little dropped a little sparkles into what ultimately was yeah. Lindsey DeFalco: the AI sparkle. Bob Moore: Ultimately was like a plug in, yeah. Lindsey DeFalco: Yeah. but yeah, I think one of the like something else I just mentioned is like d like does it matter? And like when does it actually matter? And that I think is an interesting question, 'cause like, okay, you have an agent that's barely, you know, a capital A agent, like, who cares? Bob Moore: Yeah. Mm-hmm. Lindsey DeFalco: and I think that's like super dependent on what your business is, what you're trying to achieve, you know, again, who your customers are, why you wanna say agent. If there's like a strategic reason you wanna say that you have agents, then Bob Moore: Yeah. Lindsey DeFalco: call it an agent. but we're in this funny time period where there's like a full spread of the actual maturity of these different lowercase and capital A agents. Bob Moore: Yeah, and that is like this getting back to agent fatigue, I think this is where we've come full circle because there's like a if I hear from a platform like, hey, we now have agents or use our agents, to me it's a turn off. Cause like when I hear that, I don't need another freaking agent. I got my agent. Like unless it's a very, very narrow thing that I'm gonna dispatch With an audience who's not me, and it's not internal, but it's like externally solving for something. The reality is that like I can spin up agents all day, it's not all that difficult. what I really need is like harnesses around these things that kind of help with the the context and the conditioning and the instructions that allow them to perform, or infrastructure in which to place them. Which I'm pretty locked in on already because I have my system of record platforms that that know what they know. And also, you know, we're a clawed shop here and we've got our our system of record like core horizontal internal usage agent. that that's gonna be what we we build anything for our internal use on, right? So Lindsey DeFalco: Mm. Bob Moore: but it's kinda like I, you know, Company X that suddenly has an agent for me, that's kind of on the fringes of what we you know, what we get value out of or helps us run our business is like actively uninteresting. and not something I Lindsey DeFalco: Yeah. I think it's Bob Moore: I care to look at. Lindsey DeFalco: Yeah, and like to your point, I think it's actually complicating integration and architecture discussions. So like to give an example, I had a I had a conversation with a large enterprise client talking about using cross beam data alongside other you know other data in a go-to-market workflow. So like cool, standard, excellent cross beam use case. but the conversation Bob Moore: Mm-hmm. Lindsey DeFalco: was okay, well we're doing this, we want to do use like this agent orchestration tool to do it, so we need access to your crossbeam agent. And I was like, what cross beam agent? We don't have a crossbeam agent. and then the conversation turned into okay, like what do you actually need? You need our data. and like the implementation. that you are proposing requires that it's in an agent, but you can create you can either have it reference in MCP instead or you can create you know an agent shell if you need to use Bob Moore: Mm. Lindsey DeFalco: agent to agent like the agent to agent protocol, for example. And so it's like if a company says they have an agent, then immediately you're like, okay, well is that how I connect to get my data? Like do I have to use A? Can I use your MCP? And there's just this whole like alternate discussion about how you're actually gonna architect and implement things if agents come into the mix. And I don't think companies have gotten to like cross org agents talking to each other yet at scale. but, you know, when that does happen, then this is just gonna be like another component of, you know, how you wanna pull in data from another company and why you do it from an agent versus a an C P and then Bob Moore: Mm, mm, mm. Lindsey DeFalco: you also have another another one that's like a different layer is you have you know companies like AWS for example who their C P tool is actually starting a conversation with an agent. And so it's this like backwards I think people think about like agents like wrapping C Ps with, you know, reasoning. Bob Moore: Mm-hmm. Lindsey DeFalco: But they did it this other way where like their MCP actually starts a conversation with their agent and then like returns it through the MCP. So I I think I've said this before, but it's a very interesting time where I feel like people are talking about implementation and architecture significantly more out in the open than any other time I've been in this in this field. Bob Moore: Mm-hmm. Yeah, that's true. I I can't recall you know what it is? It's product it it's product people talking about it, is what's interesting. Like yeah, well we are too yeah, yeah, Lindsey DeFalco: you guys aren't all talking about it too. Bob Moore: yeah. I very, very true, very true. I guess there's like I'm trying to think about where and in what context would I have found a discussion about like a previous generation piece of Varkat like It's like, I really need a Redis c it's two thousand and eleven and I really need a Redis cache to, you know, deal with some kind of some kind Lindsey DeFalco: Yeah. Bob Moore: of issue happening between my my database and my front end. That's an engineering conversation. like that piece of the stack, like I might find some some discussion about it on Stack Overflow and it kinda dies there. But yeah, this is like the these these these infrastructure components are becoming kind of like mainstream like technical decisions for anybody inside of a inside of a technology company. Lindsey DeFalco: And it's like it shifted the discussion too. I I I have never encountered so many like business people that I've had to even start discussing like, I don't know, what's your architecture and implementation. and I think that I mean in general AI has shifted a lot of very technical things over to you know, more knowledge workers and and traditionally business oriented folks and that's kind of Bob Moore: Mm-hmm. Lindsey DeFalco: the same thing I'm seeing which is interesting 'cause I think there's a lot of there's like a weird blend of business and technical knowledge. that doesn't actually fully live in any one team or person's head. Bob Moore: So I'm having this this flashback to back in my old like computer science courses in college where I swear we spent like half of our time in in my first computer science class, we were learning C and we spent half of our time on these like malloc functions, these memory allocation functions. So it's like, you want to store a var variable? Allocate memory. Like the the first Lindsey DeFalco: Mm-hmm. Bob Moore: thing you need to do is is set aside some memory and it's gonna be allocated and then we're gonna place this val value in the memory and then Later on, we're gonna delete the allocated memory and everything was all about like the the extreme efficiency of being able to store things, largely grown out of the fact that when C arose as a as a language, efficiency of storage was really important because storage was so incredibly expensive and RAM was so incredibly expensive, especially. and and these applications, you know, were like sitting in RAM while executing. So that was in that was ingrained in me. And get your bingo cards out. Cause at RJ Metrics, Lindsey DeFalco: Mm-hmm. Bob Moore: we had this thing that I still get made fun of to this day, which was we needed to store like calendar dates. Like if you had a chart and it was like, it's gonna run from this day to that day, and you'd physically store the date. There was like a start date and an end date in the database. But then we wanted to allow for, hey, what if you want to view like the last six weeks, like a rolling date range? And rather than add an extra column to the database to cover that scenario or have have a more sophisticated thing that require more storage, I wrote this thing where if the date stored in the date time value in my SQL was my birthday, like date and year, then That signaled to the application that the hours, months, and days represented the quantity and the units of what the rolling date range was. So if the date stored as like the starting date for the chart was 112383, there goes my ATM code. Just kidding, that's not my Lindsey DeFalco: Cause you have to Bob Moore: ATM code. so was was my birthday, then the hours would represent like One represented days, two represented weeks, three represented months, four represented years. So whatever value was there. Lindsey DeFalco: This is ridiculous. Bob Moore: And then the minutes and hours was like base 60 encoded. So you could have a maximum of 3,600. So like if I wanted to have a chart that said, show me the last two weeks, then it would be the date, the starting date of the chart was 112383, and then it was at two hours, zero minutes and three seconds. So it's like rolling date in weeks three. And then the system knew to show the chart for the last three weeks because of that. And if anybody ever tried to start a chart on the actual day, 112383, they would have had some kind of very weird behavior happen. But it never happened. Lindsey DeFalco: Mm-hmm. Bob Moore: so why did I do this? Because I learned in those early computer science classes that like storage and efficiency matters. And like I hate you can ask the trades team. I'm always like, no new database columns. Like stop adding database columns as a solution to your problem that you haven't been able to organize your data right. and that is a pepee of mine to this day. But why do I bring all this up? is anyone still listening? Lindsey DeFalco: Yes, why Bob Moore: Why do I bring all this up? Because What I learned later and what I've since observed is because storage has gotten so cheap and RAM has gotten so incredibly cheap that no one programs like this anymore. And the amount of data that gets stored that no one uses or no one needs, the width of tables, the complexity of objects that sit inside of those tables where we've got like JSON sitting inside of database fields and these incredibly like The number of megabytes that your s computer needs to download to go to Apple's homepage just to see a picture of an iPhone, the world is wildly different. And storage has gotten cheaper, bandwidth has gotten cheaper, et cetera. People don't build with regard for storage and bandwidth and memory use. They just assume that it's effectively infinite and they do all kinds of really interesting and amazing things, but they also just save a bunch of time by not even thinking about it as as a constraint. So here we are talking about all this token minimizing and token mining. And I get a little bit of this like spidey sense of like, is all of this like, hey, we should have these closed loops and determinism and everything else. Are we a bunch of old fogies Lindsey DeFalco: Hm. Bob Moore: acting like like we're running malloc functions on LLM tokens? And the reality is the cost of tokens is going to drop so radically, and the consumption of tokens is gonna grow so exponentially. That all these little efficiencies we're talking about aren't even worth the time. And maybe we should just be fire hosing all this stuff into LLMs and having them do literally everything. I can't help but have that thought. I can't help but have Lindsey DeFalco: Yeah. Yeah. Yeah, I mean that's a great great point. and I'm immediately like then there's the whole environmental impact and all of the ethical impact of fire hosing everything into LLMs. mean hey, again, It's yeah, it's hard to know what you're like rooted in from your past experience versus where things are going. It's very possible in six months that everyone's like we have replaced every deterministic model with something that we just feed into the L L to figure out 'cause tokens are free, basically. Bob Moore: Yeah. it does, you know, and and there are some things where determinism is is a feature, not a bug, right? I mean, like if you're doing math, if it like yeah, Lindsey DeFalco: Yeah, like security. Bob Moore: there are things where it's just like the wrong answer to the problem. And I think what's what's interesting is in low stakes situations where determinism is actually a better solution. I think we're gonna have a lot of people throwing stuff into LLMs anyway, not because it's Lindsey DeFalco: Mm. Bob Moore: cheaper, not because it's better and more reliable, because it's what they know how to do. Because we have this whole generation of like business people and non-technical, non-engineering people that do not know how to even write an Excel formula, have never done a V Lindsey DeFalco: Mm. Bob Moore: lookup in their lives, have never, you know, like Lindsey DeFalco: Lucky. Bob Moore: thought about or organized any kind of database, even in like Notion or or something. Lindsey DeFalco: Mm. Bob Moore: but they know how to use Claude to accomplish something. And Lindsey DeFalco: Mm. Bob Moore: like this is the I I bet I would love to get the data on this because I'm sure they have it and it's probably jaw dropping. But how many people put simple arithmetic into Chat GPT right now to get answers out? Like, Lindsey DeFalco: yeah. Bob Moore: don't open their calculator app, Lindsey DeFalco: Mm-hmm. Bob Moore: go to the L L and put a math problem in. Lindsey DeFalco: Yeah. Bob Moore: I bet it's I bet it's insanely high. I bet it's tens of millions of queries a day. Lindsey DeFalco: Mm-hmm. And there are like I mean, there are things I'll find myself asking ChatGPT and I'm like, you could Google this. Bob Moore: Mm, mm-hmm. Lindsey DeFalco: and why I don't will be for a variety of reasons. but sometimes it's like pure laziness, 'cause I don't wanna even have to think if a result I'm looking at is the right result. And then like other times it's because I think that AI is still bad at things like recommending stuff. we talked about my bags. It gets weird when you're doing like niche recommendations, Bob Moore: Mm-hmm. Lindsey DeFalco: but and then I I don't wanna have to type all the stuff about don't make s don't make anything up. Only pull, you know, direct reviews and quotes and all that all that other crap that surrounds it. but it is interesting, like I don't know, as a personal anecdote, I do try and still use my brain and be aware of when I'm using AI to be lazy for something that's like intellectual. Bob Moore: Mm-hmm. Lindsey DeFalco: and yeah, I have no idea. I mean, I'll sit next to kids at like a restaurant who are sitting there talking to AI the whole time and I'm like, gosh I don't know what Bob Moore: yeah. Lindsey DeFalco: it's gonna do for people's creativity and ability to reason themselves. It's kinda kinda crazy. Bob Moore: Yeah, it's that is a that is a very real thing. yeah, if we're in a simulation, it's like we've been placed in this one in this moment, to have had the benefit of living in a world where this stuff doesn't exist. but now we get to kind of realize that it does exist. And then at some point we'll all put the crazy world VR helmets on and snap back into reality. Lindsey DeFalco: Yeah, that's your you already did your ad for the what are this called Metaglasses last week, so we're Bob Moore: yeah, that's right. Yeah, I'm all out of I now I have to make up products that don't exist yet, to plug them. Lindsey DeFalco: So let's see, we covered agents. I would love in again, like three to six months, what are people actually calling an agent if we revisit that? Bob Moore: Mm-hmm. Lindsey DeFalco: what are all these things that have rolled out? Have they like turned into things? It 'cause even if you think to use the copilot example, like I mean I guess sure there's still copilots, but like are there? I don't like I don't really hear about them. pretty much you hear about like Microsoft Copilot, but that's like a whole, you know, product now. Bob Moore: Yeah, it's gotten it's gotten totally it's gone out of fashion. It's it's at the yeah, Lindsey DeFalco: Yeah. Yeah. Bob Moore: it's out of it's you know, you don't wear white after Labor Day and you don't call Lindsey DeFalco: Yeah Bob Moore: things a copilot after Labor Day twenty twenty six. Yeah. Lindsey DeFalco: There you go. yeah, and I'm curious, like I think I think that we will end up with more strict standards around this stuff as more people are building them into their like larger processes and there's more architecture around them. because you'll need to say like, okay, cool, you have an agent, but like is it an agent? and you know, we've also, even at Crossbeam seen some crazy like security review requests come in. If you say you have an agent, all of a sudden the diligence around like what it has access to, you know, the bounds of what it can reason on, Bob Moore: god, sure, yeah. Lindsey DeFalco: you know, it gets it gets really really complex really fast. So yeah, I'm I'm really curious to see the evolution of all of that. And I think the other the other big takeaway from me here is like are we just old old farts who are being holding on to deterministic workflows? Bob Moore: You can stop right there, yeah. Lindsey DeFalco: And anyone who's working at like an AI frontier model company who listens to this would be like pfft We do everything with LLMs. Bob Moore: No, look, and you know what? maybe, but you know what else I'll say is like I I am constantly I'm I'm so incredibly you know impressed by the business organizations that exist behind all these AI native companies and the rate of growth that these businesses experience. And like I'll I'll say good things all day about how I admire them. But the reality is I have met and interacted with many people that work at every single one of these companies across the full gamut and If you think there's not Excel sheets with V lookups powering really powerful stuff and Lindsey DeFalco: Ha ha Bob Moore: PDF documents flying around and like things that feel like never within a million miles should they exist in this magical organization that uses LLMs for everything and it all just like works in a profound way and they have two employees, you're wrong. There's a lot of humans and there's a lot of process and a lot of it is not is not solved by the fact that they have unlimited access to the world's greatest models. now look, the research divisions and everything else, like whew, otherworldly. But I think there's also a business inside of these businesses. and Lindsey DeFalco: Yep. Bob Moore: those businesses are powered by people, and those people have things they're familiar with, and they are probably really sharp and really good at leaning on LLMs, but Yeah, they haven't replaced the the the need to have thousands or tens of thousands of employees, which is the scale that that these model companies are at now. And it's not because of the small teams of researchers that are driving the improvements in the models. It's because of the go to market organizations and the the revenue scale that's being driven. Lindsey DeFalco: Yeah, it's been watching the model companies grow has been fascinating because even all of the, you know, fear mongering articles about all these jobs, you know, the entire job sector going away. if you just look at these companies, like you said, like they're they're still operating businesses. They still have people within all these different functions. And like the amount of people they need in different functions, is is slightly different sometimes, but like they still have sales teams, they still have C Bob Moore: Mm-hmm. Lindsey DeFalco: S teams, they have engineers, they have product people, like they have all of it. And so it's interesting 'cause like I think all of that stuff, all of the the scary stuff would feel more scary if there were these companies that were truly operating radically differently. and I still don't think we've seen that at like a you know, an enterprise level scale yet. Bob Moore: Yeah. yeah, I think the the things that get scary are the ones where I don't know, hacking into the nuclear launch codes doesn't require an enterprise sales force. Lindsey DeFalco: yeah. Bob Moore: you know, th this is where it gets kind of like engineering bioweapons doesn't require figuring out how to reconcile your CRM data against your marketing automation systems data and it it's actually kind of interesting like the classes of problems that are hard because they're hard versus the class of problems that are hard because they're annoying, or they're born out of, you know, weird legacy artifacts and how people try to make money and build businesses. I think our heads are in this this business realm, so much that where Lindsey DeFalco: Yeah. True. Bob Moore: we're at the frontier I think it gets terrifying Maybe has more to do with the stuff that's novel than the way that they solve against these these annoyances we've created for ourselves in in the business world. I need to adapt Lindsey DeFalco: Yeah, like Bob Moore: to. Lindsey DeFalco: w we're talking about the business of the technology, but like just the technology itself is and here we are again, towing the line of things that I do not feel qualified to comment on from an from an ethical standpoint, which I think is an impossible task when you talk about AI in a podcast. So Yeah. The thing I w Yeah. Bob Moore: That is the truth. Well, hey, I have a challenge that I wanna issue to our listeners. wait, you you I I'm I'm making a full pivot. It looked like you had one last point. You got it? Lindsey DeFalco: The I was gonna say something that may or may not be motivating to people, which is like the the ultimate thing that I keep holding on to right now is like no matter where you are in your AI journey or where your company is, like you are vastly ahead of someone and you are vastly behind someone else. And that it's v Bob Moore: It's a great point. Lindsey DeFalco: it's very important to be talking to both of those people. And the way that I've been trying to approach it is like find the people where I feel incredibly dumb having a conversation with them and learn as much as I can. And then find the people that are are you know significantly behind where either I am personally or where we are as a company and then try and help them. And I think that's like the the way that I look at it is like we're we're very much on like there's just such a huge spread right now of like skills, expertise, knowledge across all these different things. And it's moving so fast it's impossible to keep up. and so yeah that's just a a little motivational note if anyone ever feels behind in their knowledge about any of this stuff. Bob Moore: Yeah, it's so great. That's great advice. I think that's tr probably like true in life in general. it it Lindsey DeFalco: Yeah. Yeah. Bob Moore: is an interesting thing over time what'll happen to anybody who who kind of lives that way and pursues knowledge that way is like you reach a point where someone who you used to feel really dumb talking to them suddenly feels like they're at the e other end of the spectrum, of of the people that that maybe you feel like are learning from you. and that's that's what growth looks like. And everybody's an expert in their own domain. So it's it's cool to it's cool to pursue it that way. well hey here's the here's the challenge I was gonna issue to our listeners. It's not a challenge at all. it's a little bit of a a test that I'm running because we're getting tons of downloads of this podcast. It's really exciting. We're talking about different formats, maybe starting to bring guests on. We're gonna start promoting it a lot more. But I really am curious Is anybody out there listening? Like anybody at all? Like if you are out there and you're not an employee of ours or literally one of us and you're listening to this podcast right now, I want you to email me, bobacrossbeam.com, and just say, hey, I listened and let me know who you are. That's it. And I will send you an awesome crossbeam swag pack, including a signed copy of the ecosystem like growth book and Pam plushie and whatever you want from our swag cabinet. And I don't know, some other surprise that I won't mention here. That'll be fun. just shoot me an email, that's all you gotta do. and I'm really interested to see what comes in, who's out there listening. The the means of collecting info on podcast listeners is very, very challenging. So we love to get a little splash of that data. So just shoot me an email and we'd love to connect. Lindsey DeFalco: Yeah, awesome. I'm super curious, like Bob said, you know, if we bring in guests and start to promote this, like which episodes have been the most interesting? What kind of stuff do people like to hear about? is it actually when we talk about ecosystems? is it when we do ads for metaglasses and weird little products that print out receipt paper? Let us know. Bob Moore: Hm, yeah. I have my answer. Lindsey DeFalco: All right, well I think that's it for today. Bob Moore: All right, cool. Well hey, great up and we will catch y'all next time here on the AI Ecosystem Podcast. Have a good one. Lindsey DeFalco: Mm bye.