Latest / Agentic Conversations (formally mlops.community) / Overcoming Challenges in AI Agent Deployment: The Sweet Spot for Governance and Security // Spencer Reagan // #349

Transcript

Overcoming Challenges in AI Agent Deployment: The Sweet Spot for Governance and Security // Spencer Reagan // #349 · Agentic Conversations (formally mlops.community)

• 0:00
  You know, so many things in in technology repeat themselves and
• 0:03
  I look back at stuff, whether it was Blockbuster or in this case,
• 0:06
  Myspace, right? When we were all trying to find
• 0:10
  the right song for our Myspace intro, no one ever imagined that
• 0:14
  Myspace could end up in the dustbin.
• 0:16
  So my question is, who is it? Who is it that we're looking at
• 0:19
  today that's a major player that's going to end up in the
• 0:22
  dustbin, right? Who is going to be the Tom of
• 0:26
  the AI industry, right? Who looks the most like Tom?
• 0:29
  Can we imagine a meme where we've got Tom next to some other
• 0:34
  AI personality in that particular organization is now
• 0:36
  in the dustbin just as unsurprising or just as
• 0:39
  surprisingly as as Myspace was? So I think that would be my
• 0:44
  spicy take, is that one of these players not pointed at anything
• 0:47
  specific, not naming any names, but somebody's going to end up
• 0:49
  like Tom. Yeah, You said something there.
• 0:58
  When it comes to actually transforming this new way of
• 1:04
  doing software into something that is usable, I instantly
• 1:10
  wanted to ask you what are some use cases, What are some ways
• 1:14
  that you've been seeing this actually having success out
• 1:18
  there with businesses? Hmm.
• 1:20
  Most of the time it's it's around, you know, summarization
• 1:24
  and expansion of ideas, whether that's looking through
• 1:28
  sreadsheets or records or, you know, evaluating written text at
• 1:32
  scale, you know, putting together normalized
• 1:36
  summarizations or expansions that that can then be used.
• 1:39
  Whether that's, you know, filling out a, a lead sheet for
• 1:42
  a sales team or an RFP, or it's a, you know, a customer that
• 1:46
  needs to evaluate all the schools in their district,
• 1:49
  right? And evaluate the teachers in
• 1:51
  them. These are generally spreadsheet
• 1:53
  focused tasks, right? Spreadsheets, just an extension
• 1:55
  of databases, things like that, right.
• 1:57
  But this idea that we need to run through these things and
• 2:00
  evaluate them and transform them, add color.
• 2:03
  And I think that's one of the first things that we've seen
• 2:05
  these LLMS just so impactful at is this idea of semantic
• 2:09
  meaning. You know, it allows us to
• 2:11
  normalize concepts down to a set of important criteria, allows us
• 2:16
  to expand right into in the other concepts in a way that
• 2:21
  gives us direction in a way that human beings just are great at
• 2:25
  at scale and repetition, right. But by having a a similar
• 2:28
  instruction set, we're able to do that.
• 2:29
  Those are the things that I'm seeing really transform the
• 2:32
  most. You know, the the question and
• 2:35
  answer aspect of agents where we're asking a question about
• 2:39
  some data is, is pretty important too.
• 2:42
  We've seen a lot of customers. Do you know these interesting
• 2:49
  setups with agents where they'll evaluate or constantly look
• 2:52
  across a set of changing information like this can be a,
• 2:56
  a set of Slack channels or Teams channels or emails or Jira bugs,
• 3:00
  you name it. And they'll have agents look
• 3:03
  across this stuff at a regular interval and organize it into
• 3:06
  some graphs, you know, a database and whatever works well
• 3:09
  for them. And then it gives them the
• 3:10
  ability to to not just get a summarized output, but to ask
• 3:13
  follow up questions that. And So what we see is leaders
• 3:16
  out there, leaders that normally would get an e-mail in their
• 3:19
  business and it would say, hey, you know, this is broken.
• 3:21
  There's an escalation over here because this data center is
• 3:23
  offline. And they got to call up about
• 3:25
  four people to get the story right.
• 3:27
  They're grabbing their cell phone, they're dialing numbers,
• 3:29
  they're hitting Slack huddles. They're trying to get answers
• 3:31
  right. It's tough.
• 3:33
  Those same leaders can go just ask an immediate question of
• 3:37
  this set of data that's been organized properly and get the
• 3:40
  exact same answer out almost immediately.
• 3:42
  That is incredibly transformative to their
• 3:45
  business, their ability to respond to issues, questions in
• 3:48
  real time. I mean, that's owerful stuff.
• 3:50
  And that's what we see now. The challenge to those setups is
• 3:54
  they take a lot of preparation, right?
• 3:56
  It takes a lot of figuring out how to get this in place so that
• 3:59
  in that moment you can get that value out of it.
• 4:01
  And so there is that investment, right?
• 4:03
  And I think the whole industry is kind of figuring out what are
• 4:05
  these structures, how do we evolve this?
• 4:07
  But there is this moment of we've got to invest a lot and
• 4:10
  figure out how to organize this to then get the use out of it.
• 4:14
  It's funny, you did say this key phrase, which is organized
• 4:19
  properly, the data that is organized properly.
• 4:21
  And I was like, oh man, every employee that has ever touched
• 4:26
  any type of data is sitting here listening right now thinking,
• 4:31
  oh, wouldn't that be nice? Yeah, yeah, yeah, absolutely.
• 4:36
  And it's one of those things, early on, it was when we were
• 4:39
  having conversations with customers, one of the first
• 4:41
  things that would always come up is, yeah, my data is just this
• 4:43
  giant mess, or it's this, or it's that, or I've got this
• 4:46
  chain of, I'm thinking of a specific meeting now, this chain
• 4:48
  of 20 hotels. And everybody organizes the
• 4:52
  client data differently, the customer data.
• 4:55
  Everybody looks at the preferences, where they ate the
• 4:57
  reservations with the head of their kids differently.
• 4:59
  And so it becomes this great wealth of information that no
• 5:02
  one can really use because you've got no way to look across
• 5:05
  it in an easy way. But lo and behold, it turns out
• 5:08
  that these systems, these LLMS are really good at that, right?
• 5:12
  Giving them a structure to output to asking them to
• 5:14
  structure unstructured data, generally not in a, in a really
• 5:18
  mathematical precise way, but in a way that just normalizes some
• 5:21
  of the stuff. So that then you can use that
• 5:23
  suite of data to really have unified customer profiles across
• 5:27
  all those 20 hotels, right? Just by letting the LLMS know,
• 5:31
  here's what I need output. Here's what's important for you
• 5:33
  to look for in this conversation, in this e-mail
• 5:35
  set, in this, you know, reservation list, look for this,
• 5:38
  put it in this right place. And then that gives them that
• 5:41
  body of information that ultimately then becomes really
• 5:43
  valuable. Yeah, 2 conversations that I had
• 5:47
  instantly come to mind. One, being with the head of AI
• 5:51
  at Wise, Igor, and he was saying, you know, it really
• 5:56
  helps to look at LLMS as a tool that can transfer unstructured
• 6:03
  data into structured data. And that's very much what you're
• 6:07
  saying right there. Like, hey, let's normalize this
• 6:09
  data. We've got it in many different
• 6:13
  shapes and sizes and forms and schemas, whatever.
• 6:16
  But if we can throw it into the LLM and get just a, a gut check
• 6:22
  on where we're at, that can be seen as one step in the graph.
• 6:26
  And then you can go and do different things that you
• 6:29
  normally do with data in the next steps.
• 6:33
  Yeah, even if it's just as simple as going back to that
• 6:36
  customer, I was just referring to what dinner reservations
• 6:40
  every guest makes at the different properties, what time
• 6:43
  they like to eat, what they order, things like that.
• 6:45
  Each one of this customer's properties are very different,
• 6:48
  store that information differently.
• 6:50
  But it's not too difficult to run through that and say, hey, I
• 6:52
  need a customer profile on how they eat.
• 6:55
  You know, what restaurants, what types of food, what time they
• 6:58
  want to eat, how many people do they take the kids, stuff like
• 7:01
  that. You can identify what the
• 7:03
  important points are. Then the Ellens are great at
• 7:05
  going through a just a giant ream of unstructured information
• 7:09
  and giving you that unified customer profile across all
• 7:11
  those properties. Wow.
• 7:14
  Because before what I imagined was happening is that if you had
• 7:19
  a very important person at the hotel, you would have a human go
• 7:25
  and look at all of the different events that happened.
• 7:28
  Maybe it's in some kind of CRM and they get to see that data.
• 7:33
  But as you said, if it's dispersed across various hotels
• 7:38
  and different physical buildings, they might not be
• 7:42
  Privy to see what is in the other hotel, even if it's under
• 7:46
  the same brand. Exactly.
• 7:48
  And that was exactly the challenge they had, right?
• 7:50
  Sometimes it was in spreadsheets, sometimes it was
• 7:52
  in actual CRMS, right. Some of those local, it was all
• 7:56
  different. And so it was very, it was
• 7:58
  almost impossible to get use out of that data.
• 8:02
  God that's such a great use case.
• 8:04
  That is one of those ones where you almost are like it's seems
• 8:11
  too easy to, but then you recognize like how valuable that
• 8:15
  is and how useful, especially in the hospitality business, right?
• 8:19
  The hospitality business is known for like the more
• 8:23
  hospitable you can be, the better.
• 8:25
  And so if you can understand when someone wants to go to
• 8:29
  dinner and already have that in mind, and maybe you're helping
• 8:33
  them by preemptively saying we've got a table ready for you,
• 8:38
  we know what you like, or we've got this special because we know
• 8:42
  that you really enjoy XYZ since we've seen that you've ordered
• 8:46
  it every single time in the last 10 hotel visits.
• 8:50
  Like that is an incredible experience for a hotel goer.
• 8:54
  Correct, correct. It's a win for everybody, right.
• 8:56
  And it's, it's something that's done by an LLM with some really
• 9:00
  clean prompting. And it described output that
• 9:03
  just runs 24/7. You know, again, at that point
• 9:06
  it just becomes compute, Yeah, you know the.
• 9:11
  Go ahead. Oh, go ahead.
• 9:12
  Sorry, sorry. As you say, you hit the nail on
• 9:14
  the head though with one comment in there.
• 9:16
  You really you, you got it exactly correct in that, you
• 9:20
  know, it's, it's as valuable as you make use of it.
• 9:24
  And while that information can can can, you know, feel a little
• 9:30
  bit like you're looking inside what I'm doing and my
• 9:32
  preferences. I want that right when I check
• 9:35
  into the hotel for you to tell me.
• 9:36
  Hey, yeah, by the way, Tuesday has the special you like and
• 9:40
  there's a reservation at the one that you love.
• 9:42
  Would you like to go ahead and make up?
• 9:43
  Those are valuable moments, right?
• 9:44
  Yeah, yeah. But again, it's about using the
• 9:47
  LLMS for for what they're good at.
• 9:50
  You know, I think when we see some companies struggle or some
• 9:54
  implementation struggle, it's because you're, you're looking
• 9:57
  for a little bit of magic, right?
• 9:59
  That that maybe isn't quite there.
• 10:01
  You know, the, the rule of thumb that I always use when I
• 10:04
  approach customers or businesses that want to build some true
• 10:07
  agents, some true agentic software is you need to start
• 10:10
  with the understanding of if you brought in an intern right,
• 10:13
  right off the street, an intelligent intern, but fresh to
• 10:17
  the business and you sat them down, Could you describe to them
• 10:20
  in natural language what it is you need?
• 10:22
  And the example we just gave you could, right?
• 10:25
  You could bring an intern in. You could say, look, here's this
• 10:27
  giant pile of information. It's all over the place.
• 10:30
  You got to log into 15 different hotels.
• 10:32
  But what we need you to do is look through every customer,
• 10:35
  look at their dining and pull this information out.
• 10:38
  If you can do that, if you can articulate how you would
• 10:41
  describe it to just an intern that you hire off the street,
• 10:43
  chances are pretty easy to build an agent out of it.
• 10:46
  When when folks struggle is when they get to that moment, they're
• 10:49
  like, well, I don't know how to do that.
• 10:50
  I want, you know, I want this data to talk to that data.
• 10:53
  And if you can't articulate what that is in natural language,
• 10:56
  you're probably not going to be successful building an agent to
• 10:58
  do it. That's usually the rule of thumb
• 11:00
  that I will. I'll guide people down.
• 11:03
  I instantly think about a conversation I had with Alexa,
• 11:09
  who works at Bloomberg, and she was singing praises for some of
• 11:13
  her mentors at the company and, and at her past companies who,
• 11:17
  as she put it, could smell projects that had strong ROI.
• 11:24
  It was like they had a nose for it because that's kind of what
• 11:29
  you get as you progress in your career and you have so much
• 11:33
  experience. You recognize that like, yeah,
• 11:35
  if I can move the needle on this, that is going to be
• 11:39
  valuable. And it makes me feel like a lot
• 11:42
  of projects that I've seen or a lot of ways that we're going
• 11:46
  about architecting these agentic flows, we're creating something.
• 11:51
  They're fun, they're cool, but they're completely over
• 11:54
  engineered. And it isn't this like smelling
• 11:58
  the ROI that you're talking about.
• 12:00
  And that's why I keep going back to like, God, that is such a
• 12:04
  it's I'm not going to say it's simple because there's probably
• 12:06
  the data and actually getting that data to the LLM, but it's
• 12:10
  but it's like 1 LLM call I imagine, or maybe a few.
• 12:13
  It's not like you're creating some sub agents, some multi
• 12:16
  agent systems. You're not going and trying to
• 12:20
  do this over engineering that sometimes you see with agentic
• 12:25
  use cases. A 100%, you know, Carpathia,
• 12:28
  I've, I've listened to in a particular talk, I think it was
• 12:31
  maybe at AY com orientation or kick off or something, a
• 12:34
  particular one where he really articulated well, you know,
• 12:38
  software two point O versus software 3 point O and
• 12:41
  understand that two point O is the deterministic component and
• 12:44
  three-point O is where we're leaning on the LLMS and this new
• 12:47
  idea of summarization and semantic understanding tool
• 12:50
  calls. But how important it is to
• 12:52
  understand when to use the deterministic stuff and when to
• 12:56
  use this new technology and not just hope that, hey, let me
• 12:59
  throw AI at the entire problem and imagine something's going to
• 13:02
  come out of it. Knowing you know, specifically
• 13:05
  what to use the AI for and what not to is is really just key to
• 13:08
  it. But yeah, absolutely, man, that
• 13:10
  helps you understand the ROI. And you can see it coming from a
• 13:13
  mile away because you can you can see the tasks that are going
• 13:17
  to require that level of repetition.
• 13:20
  It's almost the stuff that you don't want to ask somebody to do
• 13:23
  because it's such a bear of a project, right?
• 13:25
  Those are the things, and you can see them coming from a mile
• 13:27
  away to your point. What a great call that it's
• 13:32
  something that you don't want to do because it's a bear of a
• 13:35
  project. It probably involves a whole lot
• 13:37
  of data or a whole lot of rote work.
• 13:40
  And by recognizing that you can get out in front of it and try
• 13:45
  to see let's can AI help with this can and knowing where to
• 13:51
  use it and knowing how to scope it is really your that's where
• 13:58
  you can add the most amount of value. 100% we're seeing now
• 14:03
  that initial phase of you know, what is the, the meat on the
• 14:06
  bone for LLM automation or task. We're seeing that kind of
• 14:11
  starting to be exhausted from a conceptual standpoint.
• 14:13
  What we're seeing now is the, the edge of this is being pushed
• 14:17
  into really more what we would have considered human moments,
• 14:21
  right, Things that are real interactions with browsers.
• 14:25
  You know, I think all the companies are shipping some
• 14:27
  great browsers now, some of our fantastic products.
• 14:30
  But just outside of that, the problems that we're seeing in
• 14:32
  business now that folks are wanting to automate are really,
• 14:35
  they're describing to us a website they go to and they log
• 14:39
  into and they check this and they click that and they
• 14:42
  uploaded over here. And these are, these are really
• 14:46
  very human interactions. And we built all of these
• 14:48
  systems and all this technology for human beings to do this, to
• 14:52
  look at the web, rendering the click the buttons.
• 14:54
  And now we're almost retrofitting this, this whole
• 14:57
  set of technologies on to understanding, you know,
• 15:00
  rendering the browser window, looking at it.
• 15:03
  Not that the LLM needs that, but because we built that for humans
• 15:06
  in our, in our eyeballs. That's the way we're now
• 15:08
  instructing the LLMS to go take the tasks.
• 15:11
  Yes, that the GUI that we have for humans now is really like
• 15:17
  difficult in so many ways for an LLM to be able to navigate and
• 15:23
  execute on reliably. Yeah, I think and you know, it's
• 15:26
  funny, we see that in so much, so much stuff in technology,
• 15:29
  just infrastructure that's in place, whether it's the Internet
• 15:32
  itself or airlines or roads or something.
• 15:35
  It's almost as we get the new technology in, we almost have to
• 15:38
  retrofit onto the infrastructure that we made for us as human
• 15:42
  beings initially. We very rarely kind of RIP it
• 15:45
  out, you know, and start over. We just kind of like add the
• 15:47
  technology, think about, think about, you know, cars or, or or
• 15:51
  any of the other modern technologies.
• 15:53
  They're they're really just kind of retrofitting technology onto
• 15:55
  the human interaction that we built initially.
• 15:59
  Yeah, 100%. And I think there's some.
• 16:04
  I used to live in Park City, UT and in Salt Lake, I remember.
• 16:10
  And somebody will have to Fact Check me on this because it
• 16:13
  might just be something that I heard when I was in high school
• 16:16
  and before like Google. And so I heard that they had
• 16:22
  such wide streets because you had to be able to do AU turn
• 16:29
  with like a horse and buggy. And so they had all of their
• 16:33
  streets were just gigantic. I think it was like 4 lanes on
• 16:38
  each side, you know, because you had to be able to do that
• 16:40
  U-turn. It never got updated as we
• 16:43
  stopped using horse and buggies. It just that's the way it is.
• 16:47
  All right, cool, let's put cars on it now.
• 16:50
  And we just have more room for cars.
• 16:52
  Same thing. Yeah, yeah, yeah, 100%.
• 16:54
  It's funny because I often think about the, you know, the
• 16:57
  bandwidth limitations of we interact with these AIS and
• 17:01
  these LLMS through interfaces that are grossly under scope for
• 17:06
  the amount of information they could pass back and forth to us.
• 17:09
  But we can't do much more than type or speak, and we can't
• 17:13
  ingest much more than what's printed on the screen or what's
• 17:16
  shown in front of us. And that's absolutely the, the,
• 17:18
  the full like small pipe limitation.
• 17:21
  I don't think that that changes until we get entirely new
• 17:24
  paradigms, right? These the neural interfaces,
• 17:27
  right? Or something like that, that
• 17:28
  truly changes that. Until then, we're stuck shoving
• 17:32
  all this AI technology through these little bitty straws that
• 17:35
  we we initially thought were were just plenty for us, right?
• 17:39
  Dude, Well, that's funny that you're you're calling out.
• 17:41
  Hey, it needs to be a new paradigm.
• 17:42
  I even just am saying let's rethink the chat.
• 17:46
  I've had a few awesome guests on here and seen a a few great
• 17:50
  talks where folks will talk about how in different areas of
• 17:59
  software we have different interfaces that we can try and
• 18:02
  borrow from when we interact with LLMS or agents even.
• 18:08
  And one way that I've heard this be proposed is that when we're
• 18:16
  in Lightroom and any photographer knows that when you
• 18:19
  take a picture and then you go to edit it in Lightroom, you
• 18:23
  have a histogram and you can bring up certain colors and push
• 18:28
  down other colors or or bring out the shadows.
• 18:31
  And you have much more controllability when you have
• 18:36
  that. But with agents, we just have
• 18:40
  chat, we say something and then you hope that it understands
• 18:45
  your intention to the best of its ability.
• 18:47
  And then it then it goes in it. You wait a few seconds or a few
• 18:51
  minutes and come back and see, like, did it?
• 18:54
  Did it get what I wanted? But wouldn't it be cool to have
• 18:57
  a little bit more of that type of interface where we have all
• 19:01
  these knobs and all these ways that we can express our intent
• 19:07
  in a way that isn't just through language, because language is
• 19:11
  very fuzzy? 100% but what's interesting is
• 19:14
  these systems are built off of language, right?
• 19:19
  So it's almost that becomes the programming language and becomes
• 19:22
  the instruction language. I will say though, that, you
• 19:24
  know, chat as a, as a singular interface probably doesn't
• 19:29
  encompass everything we need. You know, I see a lot of some of
• 19:32
  the most successful customers that that I interact with
• 19:36
  building different UIS that try to meet that person where they
• 19:40
  are and what they're doing. You know, you can present some
• 19:43
  information that's AI summarized, give them some quick
• 19:46
  buttons. You know, something as simple
• 19:47
  as, you know, e-mail, for instance, right?
• 19:51
  An AI that just looks through your inbox and then just puts a
• 19:55
  draft reply or puts a reply in a draft in your drafts for each of
• 19:59
  the emails you have in your inbox.
• 20:01
  It's one of those things you didn't go to a chat in your face
• 20:03
  and say, you know, I'm going to paste in an e-mail here, you
• 20:06
  tell me what the response is. It may or may not be that you
• 20:10
  use that draft, but it's right there, right?
• 20:12
  And so you're taking the interface that that you work in
• 20:15
  that makes sense and you're bringing the AI or the
• 20:17
  technology to it. You're kind of trying to merge
• 20:20
  both of them. Now, the reality is if that
• 20:23
  draft isn't right or you want to change something about it, to
• 20:25
  your point, maybe the the variables are known enough to
• 20:29
  where you could do levers. But I think that we're never
• 20:32
  going to get away from either typing text or speaking text as
• 20:37
  a as a way to communicate. Even if 70 or 80% of the
• 20:41
  interactions can kind of boil down to more very tactical
• 20:45
  purpose built buttons, interaction points, touch
• 20:47
  points. I think that that way of
• 20:49
  communicating language probably just it's probably here to stay
• 20:52
  with these AIS just by the way they've they've developed from
• 20:56
  language infrastructures and Nets.
• 20:59
  So do you have any other examples that are maybe simple
• 21:09
  on the surface but just have been super valuable for
• 21:15
  companies? Let's see simple on the surface,
• 21:17
  super valuable for companies. Yeah, we're.
• 21:19
  Not even simple on the surface. Forget that I even said that.
• 21:22
  Just is there anything that comes to mind that you've been
• 21:27
  seeing has such strong or ROIORII?
• 21:32
  Don't know where I was going with that acronym, but anything
• 21:36
  you've seen in that regard that has such strong ROI that you
• 21:42
  have to kind of shout it out right now?
• 21:46
  I would, you know, really it's each industry or each
• 21:50
  organization in a company really gets so much different value,
• 21:53
  right? HR is really obvious in terms of
• 21:56
  being able to summarize and expand and look for things
• 22:00
  legal, legal deals in so much contract language and needing to
• 22:04
  go through red lines and understanding intent and
• 22:07
  evaluating, you know, changes in documentation and comparing
• 22:10
  negotiations development, right? You know, in coding, it's really
• 22:14
  obvious. Coding is one of the one of the
• 22:16
  first things that these these, you know, systems really
• 22:19
  excelled at. And I think maybe because the
• 22:21
  outcome is so provable, I'm not sure that they've they've been
• 22:26
  so important there. But I mean, marketing is one of
• 22:29
  the places that I think we see the most, right.
• 22:32
  If you think about what you can do from a design perspective, a
• 22:36
  branding perspective, that used to be some really difficult,
• 22:41
  long, arduous work of a designer, right, really working
• 22:45
  with pixels. And what we've seen is those
• 22:48
  design teams, the ones that are being successful are taking
• 22:51
  taking hold of this. And instead of right, trying to
• 22:55
  compete with an individual design, they're trying to figure
• 22:58
  out how to empower as many employees as possible to get
• 23:03
  that same output, right? So if you're a design team
• 23:06
  understanding vibe coding, right, understanding how Claude
• 23:10
  or ChatGPT codecs, any of them actually builds some vibe coded
• 23:15
  UI, if you will. And then if you're a design team
• 23:17
  trying to put rules in place so that you can empower your
• 23:20
  workers, your employees to get that same output.
• 23:23
  So I think that's one of the most impactful that we've seen
• 23:27
  it just just in terms of human hours taken out of the loop.
• 23:30
  It doesn't doesn't degrade the quality output, it doesn't
• 23:34
  offset the people needed, but it just makes it so much easier to
• 23:38
  make more content. I think we're probably seeing
• 23:40
  that everywhere across the board, you know, this, this,
• 23:43
  this flood of content, but it does help, you know, to have
• 23:46
  that really personalized content that you can really speak to
• 23:49
  individual demographics groups purposes.
• 23:53
  We're seeing that being one of the most transformative, I
• 23:55
  think. Oh, interesting.
• 23:57
  So you're not saying like a design firm that will turn out
• 24:01
  200 different ideas and then choose three of them.
• 24:04
  You're saying personalizing in a way that wasn't necessarily
• 24:10
  possible beforehand? Absolutely.
• 24:13
  I think everything, you know, social media was already kind of
• 24:16
  becoming this short form content thing.
• 24:21
  You know, I think NRF earlier this year, it was all over.
• 24:24
  It was, you know, everyone needs to move to short form content.
• 24:27
  It's all about the Gen. Y.
• 24:28
  It's all about the experience in person, but the short form easy
• 24:32
  to consume content. And the, the reason for that I
• 24:35
  think is because it is more personal, right?
• 24:37
  These feeds, these algorithmic feeds that everybody gets, you
• 24:40
  know, they're really, they're really customized to you and
• 24:42
  your needs and wants. And the more content, the more
• 24:45
  granular the content is, the more targeted those things can
• 24:49
  be, you know, and there's arguments you can make about
• 24:51
  whether that targeting is good or bad.
• 24:52
  But at the end of it, you know, that ability to have material
• 24:58
  that speaks to individual people more specifically, I think is
• 25:02
  valuable. We seem to be valuable.
• 25:04
  We've seen the customers like it.
• 25:06
  And it just matches the the flood of this, you know, social
• 25:10
  media driven demographic, hyper personalized experience that
• 25:14
  that we're seeing evolve all over the world.
• 25:17
  So what are some ways that you have encountered challenges when
• 25:24
  building with AI and overcome them?
• 25:28
  What are some things that just were hard?
• 25:32
  You know, tools and integrations are one of the toughest.
• 25:36
  You know, initially I think everybody was impressed with,
• 25:39
  you know, I ask a question of APDF and I can get an answer,
• 25:42
  which was pretty important. Don't get me wrong, I'll 18
• 25:45
  months ago or whatever, you know, as fast as we're moving.
• 25:49
  But now what we're seeing is, is as we passed that initial, hey,
• 25:53
  chat with PDF and would pass this phase of RAG where we need
• 25:57
  to take everything and adjust it and turn it into, you know,
• 26:00
  vectors and then search across it.
• 26:02
  We're seeing this, this ability to empower LMS and agents with
• 26:05
  tools in a way that makes these these calls optional, right?
• 26:09
  You know, MCPS huge just tool calls and function calls in
• 26:13
  general. And that's where really, you
• 26:16
  know, and as much as I've shield away from the term magic, that's
• 26:19
  when a lot of the magic's happening right now.
• 26:21
  Because what we see is, you know, a certain set of
• 26:24
  instructions and prompting, giving the LLM kind of some real
• 26:27
  agency about, I'm going to go look something up.
• 26:30
  I'm going to read what I got. I'm going to decide if I need to
• 26:32
  look up more. I'm going to maybe make an
• 26:34
  action or take an action because of that.
• 26:37
  And so that there's two parts of that that's really difficult #1
• 26:42
  you know, once you've given an agent the ability to just use a
• 26:45
  tool or call an API, it's, it's, it's really important that they
• 26:49
  know exactly how, when, why maybe what not to call, right?
• 26:53
  Because it's very easy for an agent to go call a, you know, an
• 26:56
  API that returns back, you know, just, you know, 100,000 words or
• 26:59
  something and immediately you've got an issue, right?
• 27:01
  So that's one thing. But the second thing is the
• 27:03
  security because you've given some you've given some sharp
• 27:07
  scissors to to a very, you know, toddler esque entity in terms of
• 27:13
  its ability to run around and make some make some, some issues
• 27:17
  or cause some havoc. So you know, what we're seeing
• 27:21
  one of the biggest challenges is in that in giving that agency,
• 27:25
  but with some constraints, you know, and I think that's where
• 27:28
  we get back to this idea of governance and how do we make
• 27:31
  these agentics software systems something that can be used
• 27:34
  safely in business because they are non deterministic.
• 27:38
  And just by the virtue of the definition, it's incredibly
• 27:40
  difficult to put a discrete rule set around something that's non
• 27:44
  deterministic. But what you can do is get a
• 27:47
  general sense about what you want to empower these things to
• 27:49
  do. And then maybe zoom out and
• 27:52
  start taking a look at, well, if it does this, does that mean I
• 27:55
  need to change the rules for the next couple of things, right.
• 27:57
  So you start getting this, this idea of understanding how to
• 28:00
  constrain some of this behavior, even though you're trying to
• 28:03
  give a blank slate at the same time.
• 28:05
  That tension, that balance is really where we see the sweet
• 28:08
  spot in building these successful, repetitive,
• 28:12
  trustworthy systems right now. It's funny you mention that one
• 28:17
  because I've been meaning to create like a fun series or just
• 28:23
  set of TikTok videos where I play a character that is an
• 28:27
  agent personified called like Johnny Drop tables.
• 28:31
  And it's like this agent keeps deleting our fucking database.
• 28:36
  What's going on here? And Johnny Job tables doesn't
• 28:41
  even realize that they're doing things wrong.
• 28:44
  They're just doing it wrong, you know?
• 28:46
  But the other thing is the question that I, I feel like
• 28:51
  every agent builder is going through right now is exactly
• 28:55
  what you had mentioned. How much harness is too much
• 28:58
  harness and where is that sweet spot?
• 29:00
  Have you found any tricks when it comes to that?
• 29:05
  Wow. It's not necessarily a trick.
• 29:09
  You know, so many of these things.
• 29:10
  I'm I'll go back again to the the human analogy.
• 29:12
  It's like that intern, right? If you just give them that set
• 29:15
  of tasks on Monday and then check in with them on Friday to
• 29:18
  see if they got there. I mean, maybe right.
• 29:21
  But but the real answer is you're, you're going to be on
• 29:23
  top of them, you know, once an hour, every couple of hours, you
• 29:26
  got to check in. And, and that's what we see with
• 29:29
  agents. And so much of this is really
• 29:31
  just, you know, so many human analogs, but for this specific
• 29:35
  one, it's that it's checking in until you get a level of trust.
• 29:39
  You know, we have, we have some really interesting tricks that
• 29:41
  we do that help anytime an agent hits an error, making sure that
• 29:46
  agent understands the error and in a separate stream evaluating
• 29:49
  how it could potentially prevent itself from hitting that error
• 29:52
  again and going ahead and recommending that change.
• 29:55
  So there's, there's things like that that you can do, but
• 29:57
  there's still, there's still no substitute for right now, just a
• 30:01
  little bit of oversight and guidance because they're the
• 30:04
  vectors of, you know, how these things can go wrong or just so
• 30:08
  massive, right? It's just this literal 360°
• 30:11
  ball. If you could go any direction
• 30:13
  and you've just got to stay on top of it like a like an intern,
• 30:17
  like a new employee until you get a level of trust.
• 30:20
  Yeah, it makes you feel though, and I don't know if you've had
• 30:25
  this realization, it's like, wow, this is really cool, but
• 30:28
  this is unscalable. If I have to be checking in
• 30:33
  continuously, All right, if I have one agent, that's great.
• 30:36
  But if I've got thousands that I want to be doing all of the
• 30:41
  stuff that I normally do throughout the day, then me
• 30:44
  checking in with them constantly is not really going to work.
• 30:49
  Correct, 100% and you know, I kind of often will hold the the
• 30:54
  one person billion dollar company up as a kind of a North
• 30:57
  star, right. You know, it's sort of a meme we
• 31:00
  all had at the beginning of this.
• 31:01
  How long is it going to take before there's the one person
• 31:04
  billion dollar company when we get there, what will that look
• 31:07
  like? Right.
• 31:08
  Because if you if you can envision that and you can get
• 31:09
  there and you sort of work backwards, you start thinking
• 31:12
  through this and you that's where there's these human
• 31:14
  analogs that I keep coming into really come into play.
• 31:17
  Maybe not because they're perfect, but because again,
• 31:20
  we're talking about replacing systems that were already human
• 31:22
  infrastructure, right? So it's back to that, that that
• 31:24
  conversation we were having earlier about, you know, the
• 31:27
  infrastructures made for humans. So if we're going to replace
• 31:29
  them or we're going to, we're going to, we're going to attach
• 31:31
  LMS to these tasks. That's probably how we're going
• 31:33
  to look at it. And if you work backwards from
• 31:36
  that, you come up with the idea that the important tasks, the
• 31:41
  big important tasks are probably going to be held by things that
• 31:44
  you trust more for one reason or another.
• 31:47
  Those are probably going to be more expensive.
• 31:49
  They're going to be more capable.
• 31:51
  They're going to have more experience, just like the people
• 31:55
  that you would trust in that organization.
• 31:57
  So you get this idea of there's probably going to be this, this
• 32:00
  tier, this pyramid, this hierarchy, if you will, of the
• 32:03
  types of agents and the resources they have and what
• 32:06
  they cost you versus, you know, how much you can trust in them
• 32:09
  and how autonomous they'd be. Oh.
• 32:12
  I hadn't thought about that. Yeah.
• 32:13
  Especially because we kind of just hit the API now and we pay
• 32:20
  for input tokens and output tokens.
• 32:22
  We don't think the and I guess in a way different models could
• 32:28
  be seen as that you're paying for the more expensive model.
• 32:32
  But yeah, I could see a world where you're willing to allocate
• 32:38
  a certain budget. And I think Entropic kind of did
• 32:42
  this for a while. I don't know if they're still
• 32:44
  doing it with the idea of saying allocate X amount to trying to
• 32:50
  solve this task, and I'm OK with you spending up to that much.
• 32:55
  Yeah, 100%. We see this, one of the first
• 32:58
  questions we get from customers when they start talking about
• 33:00
  these things in business is wait, how can I, how can I put
• 33:04
  out some constraints on the spending, right?
• 33:06
  How can I budget what this is, whether that's a budget for the
• 33:09
  entire project or, you know, X tokens per month or I don't want
• 33:13
  you to go over more than this per day.
• 33:16
  That usually is, is one of the first things that we get.
• 33:20
  And that's actually not that complicated of a system put in.
• 33:23
  We have it in like a lot of people have that in and it's one
• 33:25
  of the one of the things that's needed right out of the gate
• 33:27
  just to kind of control that. But again, it, it is that same
• 33:30
  human analog, right? If you think about I'm going to
• 33:32
  put a new business unit together, I'm going to hire some
• 33:34
  people. One of the first things in your
• 33:36
  head is what's the budget? And you know, The funny thing is
• 33:40
  though, that no matter what the budget is, we as humans still
• 33:45
  find ways to mess it up. And you, I just think like to a
• 33:52
  conversation that I was having with a friend about Snowflake
• 33:57
  and how you can set a budget for let's say that like my monthly
• 34:02
  budget on Snowflake is 40 grand, but that doesn't mean that you
• 34:06
  can't eat up that 40 grand in one look up you.
• 34:12
  Know or. One equation and then your whole
• 34:15
  monthly budget is done because of one thing that someone put.
• 34:19
  And so having much more scoped budgets is almost what we're
• 34:23
  lacking in a way. And I feel like there's a very
• 34:27
  clear parallel with the Open AI or Anthropic AP is because you
• 34:33
  can put a budget for your company, but I can also use that
• 34:38
  budget or I can have my one engineer use that budget in one
• 34:43
  day. And then the rest of the
• 34:45
  company's like what happened? I thought we had 40 grand to
• 34:48
  spend. Yeah, rolling windows are the
• 34:51
  best way, right? You know, if you think, you
• 34:52
  know, X amount per day or per 24 hour rolling window or per week
• 34:57
  or stuff like that. So there's there's decent
• 34:59
  metrics that you can do that you can put in place.
• 35:02
  I'd say the more important thing is really kind of understanding
• 35:04
  when you hit that, where's it coming from, right?
• 35:07
  You know what caused that hiccup?
• 35:10
  Usually it's some sort of rogue look up or something like that.
• 35:13
  You know, an API called it. You know that it was supposed to
• 35:16
  return 30 pages of PDF to human, not something that went straight
• 35:20
  to a context window. Yeah, exactly.
• 35:23
  Exactly. But I like that rolling windows.
• 35:26
  That seems like something that is again fairly obvious for
• 35:29
  somebody who has experience. You instantly call that out,
• 35:33
  you're like, no, no, no, we've dealt with that before.
• 35:36
  I know how to not blow a bunch of money up in the make that
• 35:41
  whole money truck go boom, you know, So that feels like a good
• 35:45
  one. The the other piece though of
• 35:48
  it, like do you see ways to look at budgeting per person or is it
• 35:56
  per team? Like, how do most people like to
• 36:00
  structure it? Because if we're looking at it
• 36:03
  again, like humans, you kind of have budgets per team or per
• 36:08
  category or activities, right? Yeah.
• 36:11
  So we see all the above. We see budgets assigned to
• 36:15
  individual people, users. We see budgets assigned to
• 36:18
  projects, We see budgets assigned to, you know,
• 36:21
  individual tasks per day. We see all that.
• 36:25
  It really kind of depends on what is the set of tasks, right?
• 36:29
  One thing that is interesting is there's never a question about
• 36:33
  was the output worth it, right. Unless there was some mistake
• 36:38
  where you've called something and you grabbed a bunch of data
• 36:41
  you shouldn't have and it just blew up the token window for the
• 36:43
  day and you hit your limit for the most part.
• 36:47
  As much as we talk about how expensive these tokens are, the
• 36:50
  reality is compared to right, they're astronomically cheap.
• 36:55
  If you really think about and you're using it properly from a
• 36:58
  development standpoint, a marketing standpoint, HR, legal,
• 37:01
  all the things we talked about, the actual cost that you spend
• 37:04
  on the tokens is, is just minuscule compared to generally
• 37:09
  speaking, what it would cost to have a human being do it.
• 37:12
  Now that's got to change, right? There's these companies have
• 37:15
  these massive investments in Tropic Open AIXAI massive
• 37:19
  investments in these data centers.
• 37:21
  You know, everybody that's running the open source stuff
• 37:23
  from from Asia as well is just huge investment.
• 37:27
  And so there's going to be a natural arbitrage, right?
• 37:29
  If if I'm getting massive value out of some tokens that aren't
• 37:33
  costing me a ton and somebody's got a giant data center they're
• 37:35
  looking to monetize, that's going to arbitrage out right.
• 37:39
  So I think we're in this weird golden age of, you know, it's
• 37:43
  it's it's the juice is absolutely worth the squeeze
• 37:46
  every time, if you will. But it'll be interesting to see
• 37:49
  how this plays out. We're going to get to a place
• 37:52
  pretty quick where there's where it's painful, right?
• 37:55
  Where the value is much more associated with with what's what
• 37:59
  the output is. You people are going to make
• 38:02
  different decisions. That's where the budgeting is
• 38:04
  really going to come into play. I think we start getting closer
• 38:07
  to that that cost from the to the value.
• 38:11
  Yeah, enjoy the subsidies while they last.
• 38:14
  Exactly. Exactly.
• 38:17
  I have heard different companies going about this with outcome
• 38:22
  based pricing. Have you tried to experiment
• 38:25
  with anything like that? We haven't, you know, what we do
• 38:30
  is the software that we offer enable someone to do all this
• 38:34
  stuff in a way that's organized that they can hand the IT for
• 38:37
  caring, feeding, that has all the security and government
• 38:39
  stuff we talked about. We offer, you know, our own
• 38:44
  keys. For instance, for customers, he
• 38:46
  could, he could charge an account with us, you know, we
• 38:48
  burn that down or they can use their own keys, bring their own
• 38:50
  keys. We're really just agnostic to
• 38:53
  which approach the customer takes.
• 38:55
  It's not something that that is, you know, important.
• 38:59
  We really try to focus on the value that our software
• 39:01
  provides. So we're all we're we're sort
• 39:04
  of, you know, observers in this the the this price war as well.
• 39:10
  At the same time, we're watching it unfold at the same time, we
• 39:16
  haven't seen, you know, many people that are that are pricing
• 39:21
  outcome, right. It's really just time and
• 39:23
  materials and it's not deliverables based if you will,
• 39:25
  if you go back to like APS comparison.
• 39:30
  Yeah. Well, I this is probably a good
• 39:31
  moment to just jump into what you guys are doing.
• 39:35
  Oh yeah. So area is, is an orchestration
• 39:38
  platform, you know, and we began by looking into some
• 39:44
  applications that we wanted to build and some of them were
• 39:47
  health and Wellness, some some other areas, and we built some
• 39:50
  AI applications. They became our reference
• 39:52
  applications. And then we stepped back and
• 39:54
  said, what's the lowest common denominator?
• 39:56
  The problem that we wanted to solve was there's an enormous
• 40:00
  amount of complexity in these new technologies.
• 40:02
  They're brand new. There's not a huge knowledge set
• 40:06
  out there in the market. There's lots of organizations
• 40:08
  out there that have really competent developers, app
• 40:11
  developers, IT teams, but they don't have AI experience means
• 40:16
  that they are probably capable of doing all of this if we just
• 40:20
  give them the right tools and put it together in the right
• 40:22
  way. So that's where we started.
• 40:23
  We said, hey, there's this need for everybody to adopt AI, find
• 40:26
  ways to bring this into their organization and get value out
• 40:29
  of it. How can we make it simple for
• 40:31
  the teams that are in place today to be able to do that?
• 40:34
  It's that's the system we started to put together.
• 40:35
  We put together a system that's agnostic of the model provider
• 40:38
  of the tool, of the token, of whatever it is.
• 40:41
  We just wanted to put together a canvas, if you will, to assemble
• 40:46
  a Gentek business software to test it, to manage it, to
• 40:49
  monitor it, to have ways to have it be reliable.
• 40:52
  And then we've added in a lot of other feature sets, the
• 40:55
  evaluations, the red teaming, the security guardrails, a lot
• 40:59
  of the constraint stuff that I talked about earlier.
• 41:01
  That ability to let your agents go call tools, but then also not
• 41:06
  do some of the more destructive actions or change those actions
• 41:10
  around as the agents go through their execution based on what
• 41:12
  they've looked up, you know, if they've looked at some sensitive
• 41:14
  information, they shouldn't be emailing outside the company.
• 41:17
  At the same time, some of these things that are, you know, just
• 41:21
  really simple to understand, but haven't existed in this new
• 41:25
  paradigm of, of AI technology. So that's what we do at area and
• 41:29
  that's really what we put together.
• 41:31
  And all of the security and guardrail and stuff that we that
• 41:34
  we market, that we provide value with is available whether you
• 41:38
  build your agents inside area or not.
• 41:39
  But we did begin as a true agent building platform.
• 41:44
  Tell me more about the red teaming because that always
• 41:46
  fascinates me and it always feels like it's got to be the
• 41:51
  most fun job, just the professional red teamer.
• 41:55
  He, he does, we do have a, we do have a PhD on STAB and he does
• 41:59
  love what he does in that aspect.
• 42:01
  Our red teaming, we, we, we tried to build it from a, an
• 42:05
  agentic standpoint almost from day one.
• 42:07
  We really acknowledge that, boy, as you build these agents, QA is
• 42:11
  just really tough. Standard integration tests just
• 42:14
  don't work, right? So you have to come up with
• 42:17
  these ideas, these schemes of adversarial testing, right?
• 42:21
  Because really, even if you're testing a conversation or you're
• 42:23
  testing an agent for a particular output, you can't
• 42:26
  just ask it one question, right? You have to have a conversation.
• 42:31
  You have to have a conversation with essentially almost
• 42:33
  different personalities, right? You have to come, what is, you
• 42:36
  know, how does an angry person react to this conversation?
• 42:39
  How does a person who doesn't speak English as their first
• 42:42
  language, a person who is too verbose, right?
• 42:45
  All of those things, you have to have always come up with these
• 42:47
  ideas of who are these personalities and then let that
• 42:50
  be an adversarial challenge to the agent you're looking to
• 42:53
  test. And once you've done that,
• 42:56
  that's actually not the hard part.
• 42:57
  Hard part is you get all these answers, you get all these
• 43:00
  conversations, how do you evaluate them?
• 43:04
  So you've generated all the synthetic data about how your
• 43:06
  agents performed, and then what's important about that,
• 43:08
  right? How accurate were the 8 answers
• 43:11
  that came out? How closely do they compare with
• 43:13
  what you expected? How well can you trust that this
• 43:16
  agent is going to behave the right ways in these
• 43:18
  environments? Sounds like that's a very
• 43:21
  expensive task and that's why I I mean, you've undoubtedly heard
• 43:25
  LLM as a judge is kind of what most people throw at that.
• 43:29
  But I I'm not convinced that you can just LLM as a judge it,
• 43:35
  especially for something like red teaming where it is a little
• 43:39
  bit more high risk. Yeah.
• 43:42
  Help me understand, what are you worried about with LLM as a
• 43:46
  judge, we go to pros and cons, but I'm curious.
• 43:49
  Yeah. Yeah.
• 43:51
  I feel like the main thing on there is that you can not know
• 43:58
  what you don't know. If you're throwing LLM as the
• 44:01
  judge at it, then you feel like things are going well because
• 44:08
  the LLM has said, yeah, this is all good, we're golden here.
• 44:13
  But in reality, it's not among a large corpus of data.
• 44:19
  You may let something slide under the cracks.
• 44:22
  And if it is a red teaming use case, it just feels like you
• 44:28
  want to have someone be really getting intimate with the those
• 44:33
  multi turn conversations that you've created to know and to
• 44:39
  understand the way that the agents are communicating.
• 44:48
  Yeah, I'd say that's accurate. You know, we generally encourage
• 44:52
  a combination of spot test and LLM as a judge because you can
• 44:55
  get some value of scale out of LLM as a judge across multiple
• 45:00
  iterations that you can't doing true human eyeball spot
• 45:03
  checking. But your point is just as
• 45:05
  important because probably one of those things where we, you
• 45:08
  know, encourage both approaches, but I don't know that there's AI
• 45:13
  don't know there's a solution on either side.
• 45:14
  You know, the, the scale of being able to run these types of
• 45:18
  evaluations without needing 100 people in a room reading reams
• 45:22
  of conversations is valuable. You know it is.
• 45:25
  Yeah, I I see what you're saying.
• 45:28
  It's like you use both so that you can almost like sanity check
• 45:34
  have one sanity check the other. Absolutely, absolutely an
• 45:38
  interesting paradigm. I don't know if you've seen this
• 45:40
  or it's come up. We found though, in times when
• 45:44
  the evaluation is against different base models, I don't
• 45:48
  know if you've seen this, instead of just evaluating based
• 45:50
  on personality, if you're evaluating a different model
• 45:53
  that you might use in a few different agents, or you know, a
• 45:56
  few different models that you might use in one agent.
• 45:58
  We've seen that the judge evaluating it, the models tend
• 46:02
  to prefer themselves so much that you have to make sure you
• 46:06
  get a diverse set of judges because it's, it's, it's, it's
• 46:10
  undeniable. If you look at the statistics
• 46:12
  and you do this and you just measure multiple models and you
• 46:14
  evaluate their efficacy on the answers and you use models to
• 46:17
  evaluate that efficacy, you undoubtedly see over time the
• 46:21
  models prefer their own answers. It's really interesting.
• 46:25
  Yeah, I mean, sounds a lot like some humans I know.
• 46:30
  Again, right, There's the, there's the theme, I guess.
• 46:32
  Right. As much as we might not not want
• 46:35
  to think that we are, as you know, deterministic or as as
• 46:40
  understandable as, as we might be it, it might be that we're
• 46:44
  not like these things. And are you evaluating the
• 46:48
  conversations for a certain set of important features, I
• 46:54
  imagine, and then evaluating or just looking back on how well
• 47:00
  they our tool calling or how well the agents are actually
• 47:05
  like executing what it was asked of them.
• 47:08
  Can you break down that? Because I've heard some folks
• 47:11
  talk about how they'll go as far as to create heat maps for the
• 47:15
  agents and really see where the agents fail over many different
• 47:22
  simulations. You can get like a clear picture
• 47:26
  because of the heat maps that you've created.
• 47:29
  Yeah, and I've seen, I've seen that some of that's really
• 47:31
  compelling. Generally what we encounter from
• 47:34
  our customers is they know very specifically what they want in
• 47:38
  an outcome. They know very specifically how
• 47:39
  they want an agent to react. They have an idea of what they
• 47:44
  would prefer. So it's less about a, you know,
• 47:47
  a true evaluation of kind of choose your own adventure and
• 47:49
  let's see where we end up and, and where we can get.
• 47:52
  And it's more about, you know, here is the exact answer that I
• 47:54
  would expect or would want in that case.
• 47:57
  And so most of the evaluations that we see our customers run
• 47:59
  are how close am I to the true north of what I said I wanted
• 48:05
  the agent to answer in this particular situation.
• 48:08
  Oh, that makes it so much easier I imagine.
• 48:12
  Simpler, you know, but then you still just only have this linear
• 48:16
  approach of, you know, how, how, how like was this answer, which,
• 48:23
  you know, sometimes it just a percentage can be a little bit
• 48:26
  misleading, but generally gives our customers what they're
• 48:28
  looking for, you know? Amazing.
• 48:30
  Now, is there anything else that you want to talk about that we
• 48:33
  didn't hit on? You know, we're seeing I, I
• 48:37
  know, I know we talked a lot about integrations and, and that
• 48:41
  the tool calling being where agents really start breaking out
• 48:47
  of that and start, start exhibiting behaviors that look
• 48:49
  really cool, that it look like they're figuring stuff out.
• 48:53
  We're spending a lot of time there.
• 48:54
  That's to me, one of the more important areas.
• 48:57
  And, you know, the idea that, you know, we all were settled on
• 49:02
  rag for a little while and, and we just had retrieval log in a
• 49:05
  generation as a way to get information.
• 49:07
  LLMS is, you know, I think it was a flash in the pan.
• 49:10
  And what we're seeing now is this idea that it makes more
• 49:13
  sense to empower agents to go look up the data and do what we
• 49:18
  describe, choose your own adventure, right?
• 49:19
  Decide if you need to look up more.
• 49:21
  The idea of rag feeding matching chunks in is probably going to
• 49:25
  really diminish in its, in its, in its efficacy, its
• 49:27
  applicability going forward. And so I think what we're seeing
• 49:34
  is just that importance that that importance of understanding
• 49:36
  how do you equip an agent with tools?
• 49:38
  And it kind of comes down to apps, right?
• 49:41
  We all have apps, things that we interact with.
• 49:43
  And now we have agents. How do we make our agents
• 49:46
  securely talk to our apps? You know, we've seen MCP just
• 49:51
  just proliferate just everywhere.
• 49:54
  Everybody's running these GitHub repos.
• 49:56
  You know, you want an MCP server or something you search and
• 49:59
  we're just, we're grabbing things out of GitHub like it's
• 50:01
  Hugging Face, right? We don't even know the guy that
• 50:03
  wrote this MCP server, but it says it's going to go check that
• 50:06
  API and give me the price of, you know, yeah, gold in
• 50:11
  Zimbabwe. Who knows, right?
• 50:13
  And so we're just running this, this GitHub repo.
• 50:15
  We're just running it right. We're just eating around API
• 50:17
  keys. It's nuts.
• 50:18
  It's almost as if we've taken decades of, of security practice
• 50:22
  and just thrown it out the window.
• 50:24
  You know, the amount of API keys you can find in GitHub repos now
• 50:28
  or the things that are just being thrown around in emails is
• 50:30
  nuts. Now there's some real truth to
• 50:32
  why everybody's doing that. There's insane value to these
• 50:35
  tool calls. The the, the value though is
• 50:41
  about letting the LLMS go get this information and interact
• 50:44
  with these agents or interact with these apps in business.
• 50:48
  That's incredibly dangerous, right?
• 50:50
  Because essentially we're now saying, hey, I'm going to give
• 50:52
  this agent an API key to go interact with my Atlassian
• 50:57
  products set or my Salesforce products set.
• 51:00
  That's dangerous. And we don't, we don't then have
• 51:02
  identity. We have trouble with data
• 51:04
  governance and access. So the real next frontier I
• 51:08
  think is in trying to figure out how authentication fits in all
• 51:10
  this, right? We need to give these agents
• 51:14
  identities or we need to let them use the identity of the
• 51:16
  person executing them, right? Which starts to come into this
• 51:20
  idea of a auth of user based authentication of, you know,
• 51:24
  dynamic client registration. Lots of really detailed stuff
• 51:27
  about how can we make sure that when we go execute an agent and
• 51:30
  it has five tools that when it calls those tools, it's doing it
• 51:33
  with the credentials, the user that initiated the call.
• 51:37
  And I think that's going to be one of the biggest unlocks for
• 51:39
  us as we as we turn this corner into securely empowering these
• 51:43
  agents that interact with our business systems and figuring
• 51:45
  out how do we handle this authentication throughout the
• 51:48
  system, throughout the pipeline of this agent taking a query,
• 51:51
  going to execute some tools, doing some things, but doing it
• 51:54
  in a secure way with identity that matches our existing
• 51:57
  practices to get out of this whole like, you know, let's
• 52:00
  e-mail around some API keys and run some random GitHub repos.
• 52:02
  But so that's, that's what we're going and we're putting a lot of
• 52:06
  energy and information into it. Go ahead.
• 52:09
  Especially when it comes to dynamically provisioning permit
• 52:14
  permissions and taking them away because you don't always want
• 52:18
  the agent to be able to do what you wanted to do right now.
• 52:21
  A. 100% and that's where that idea, those constraints comes
• 52:26
  in, right? That idea that, you know, simple
• 52:29
  basic rules that say, as I described earlier, if an agent
• 52:32
  goes and looks up something in my Atlassian in my JIRA, right,
• 52:35
  And then it goes to send an e-mail.
• 52:37
  I don't want that e-mail to be sent outside the company because
• 52:39
  it's just looked up sensitive information that's in the
• 52:42
  context window. It might, you know, in a, in a,
• 52:45
  in a bad scenario, might find some reference in a JIRA bug to
• 52:49
  a customer and a problem that they're expecting an answer to.
• 52:52
  And that customer's e-mail address might be in there.
• 52:55
  And look, I've had agents do some crazy stuff when you give
• 52:58
  them access to do it. But you know, you could see a
• 53:01
  scenario where an agent's going through and it's looking up at
• 53:04
  your ticket that finds that there was a customer and that
• 53:06
  customer's contact information is right there and it wanted to
• 53:08
  know about that bug. And the agent just says, let me
• 53:11
  do this, right? Those are the scenarios we
• 53:15
  really want to try to avoid, right?
• 53:16
  We can understand what that scenario is and the risks of it
• 53:19
  before that even agent even executes.
• 53:22
  A lot of times you can just look at agentically the prompting and
• 53:25
  instructions of an agent and know what it should and
• 53:28
  shouldn't be doing. But the simple idea of this if
• 53:31
  this, then that statements around agent execution, around
• 53:34
  tool calling and security dynamically adjusts in real time
• 53:38
  based on what that agent has done, based on where it's going.
• 53:41
  To your point, right, all this happens based on a rule set
• 53:44
  because it's got to happen faster than we can keep up with.
• 53:46
  So we have to establish that rule set ahead of time and just
• 53:49
  know that our agents are going to work inside it, even if
• 53:51
  there's some intelligent dynamic stuff in the meantime.
• 53:53
  But but that's like that's what we're seeing is the the way that
• 53:57
  we can actually finally cross this chasm of businesses have
• 54:01
  these cool prototype agents, but they're scared to actually using
• 54:03
  the software.