Latest / Agentic Conversations (formally mlops.community) / AI Agents Should Be Treated Like Hackers

Transcript

AI Agents Should Be Treated Like Hackers · Agentic Conversations (formally mlops.community)

• 0:00
  All right, so at the beginning of April I went to New York for
• 0:03
  the MCP Dev Summit, and while I was there I got the chance to
• 0:07
  record a few podcasts with attendees that were at the
• 0:11
  event. This is one of those
• 0:12
  conversations. Hope you enjoy.
• 0:14
  I think we have to look at these agents as untrusted, maybe even
• 0:19
  adversarial. How many people are reading this
• 0:21
  code or like think about, I think events of the last month
• 0:24
  with like Anthropic in the DoD, Think about what we're doing.
• 0:27
  We're asking citizen developers to deploy untrusted agents
• 0:32
  behind the firewall. And we're frantically wiring
• 0:34
  those things up to every single system we can find.
• 0:37
  And, and every CEO is basically standing up and saying we must
• 0:41
  do this. We must do it now.
• 0:42
  We have to take more risk. We have to act like a startup.
• 0:44
  I can't think of any part of our business that isn't rapidly
• 0:48
  changing where we look at it first and foremost as what do
• 0:53
  these tools let us do? But they're only useful, They're
• 0:55
  only useful if they have the context of all of the, the
• 0:59
  systems and the services and the APIs.
• 1:01
  And that's, that's what APIs are at the end of the day.
• 1:03
  They're the, they're the front door to all of that stuff.
• 1:16
  Hi, I'm Alex Alkever with the Linux Foundation and the Agentic
• 1:19
  AI Foundation. We're here at MCP Dev Summit in
• 1:22
  New York City and I'm chatting today with Matt Deburgolis.
• 1:26
  If I said it the right way, who is the CEO of Apollo Graph QL?
• 1:29
  And forgive me if I get anything wrong, correct me, I have no
• 1:32
  ego. Thank you for joining us.
• 1:34
  Before we get started on this conversation, which is going to
• 1:37
  be about API lessons for MCPI, believe is what we were
• 1:41
  discussing and tell us a little bit about what a polygraph QL
• 1:44
  does. And then how we can chat a
• 1:47
  little more about how you're leaning into MCP, particularly
• 1:52
  since you're essentially a graph QL as a service provider and you
• 1:56
  know, we're generating everything around the API space.
• 1:58
  So MCP is a very interesting translational layer for you.
• 2:03
  Yeah, sounds good. Thanks for thanks for doing
• 2:05
  this. It's a great conference.
• 2:07
  Thanks for coming. Energy is awesome.
• 2:09
  It it you can't make this stuff up, right?
• 2:12
  It it shows there. It's crazy, literally like
• 2:14
  things are so I mean like I'm following all.
• 2:17
  So each week I do a scrape of all the protocol repos and
• 2:20
  you're just seeing things move so quickly, like, oh wait,
• 2:23
  somebody signed ACLA or oh wait, you know, they just shipped the
• 2:26
  new securities back. It's it's, it's crazy.
• 2:28
  I think what's cool about MCP is it, it reminds me of the the
• 2:33
  mash up era, right? It's it's, it's for builders,
• 2:36
  it's for making stuff. And you know, we could talk
• 2:39
  about agents are useful because they can connect to stuff.
• 2:44
  So we need an answer to that. And MCP is a big part of it.
• 2:47
  And what we've been doing it Apollo for the AP is of large
• 2:51
  companies, especially is, is we're, we're basically solving
• 2:54
  this problem of how do you, how do you take a whole bunch of
• 2:56
  different systems? And AP is that you've got, you
• 2:59
  might have thousands of them in a typical enterprise.
• 3:01
  And how do you glue those together to make a great user
• 3:04
  experience? And this got its start back when
• 3:06
  we're talking about building a mobile app or a really great
• 3:10
  JavaScript app in the browser. But it's always been about
• 3:13
  builders. It's always been about, I want
• 3:16
  to make a cool thing. I want to make something my
• 3:18
  customers are going to love. And I want to do that on top of
• 3:22
  real stuff. And and so now we can have that
• 3:26
  whole conversation again, but at an even bigger scale because
• 3:30
  we're talking about with AI being able to build so much more
• 3:35
  and not just the sort of user facing, you know, first class
• 3:40
  mobile app, but think about all the things that are happening
• 3:43
  inside the business now that we've got agents on people's
• 3:46
  desks, now that we can build software large and small really
• 3:50
  quickly. And the part that I'm just so
• 3:53
  excited about is we know how much amazing stuff is out there
• 4:00
  inside micro services or inside little pieces of SAS or inside
• 4:05
  these systems that people have built over the years.
• 4:08
  And the more of those we can connect together, the more of
• 4:10
  those we can connect to an agent.
• 4:11
  I think it's going to change so much.
• 4:13
  It's going to make businesses faster.
• 4:14
  It's going to be better customer experiences.
• 4:18
  It's real exciting. And so from your viewpoint too,
• 4:21
  because you've spent a lot of time in the API world and now
• 4:24
  you're essentially living in 2 worlds, what is the transition
• 4:29
  or? I wouldn't say transition
• 4:30
  because that implies one goes away.
• 4:32
  Ape guys don't go at all. No, they certainly don't.
• 4:35
  No. And like even with like graph
• 4:36
  QL, it was still like kind of ascending, you know.
• 4:39
  And rest is still, there's a lot more people use rest than graph,
• 4:42
  even though graph is arguably much better for AI.
• 4:46
  So if you could maybe take a stab at like, what is this
• 4:50
  translational layer look like if it works really well for what
• 4:55
  we're trying to to to build that world you're describing?
• 4:57
  Yeah, I think all these things are complimentary.
• 5:01
  With Graphical, for example, it's not about getting rid of
• 5:05
  REST AP is or whatever you had before and replacing it with
• 5:08
  Graph QL. That's not what Graph QL is.
• 5:09
  Graphical is a query language, so it's, it's something that you
• 5:13
  can layer on top of your REST AP is and it makes them better.
• 5:16
  Like graph QL introduced the idea of strong typing in AP is
• 5:21
  it's got this idea of introspection.
• 5:23
  So you can ask the graph, what, what do you have for me?
• 5:28
  What are the types? What are the fields?
• 5:30
  How do they relate? How are they used?
• 5:32
  There's all this metadata that comes along with that.
• 5:35
  And I think MCP very similarly, right?
• 5:38
  Nobody's saying you're going to throw away your underlying AP is
• 5:43
  we're talking about a new way to use them.
• 5:45
  We're talking about adapting them for the needs of an agent.
• 5:49
  But I think really the, the magic thing about MCP is the
• 5:54
  flexibility in the free form part of it.
• 5:58
  You know, we're, we used to think about AP is as these rigid
• 6:02
  things that got designed very carefully a long time ago and
• 6:05
  they never change, right? Like, but that's not the world
• 6:08
  we're in anymore. Systems can change every day,
• 6:10
  and MCP gives us a way to to talk about that and to expose
• 6:15
  those ideas to a system. Dive a little deeper on that, if
• 6:18
  you can tell, I mean, if you will, on what is
• 6:21
  architecturally, what does that start to look like if you're
• 6:23
  trying to build out something that takes advantage of the
• 6:27
  goodness of graph QL and melds it to the goodness of MCP?
• 6:31
  Yeah. So a company has been using
• 6:34
  graph QL has, you know, we call it a super graph at Apollo.
• 6:37
  They've, they've got this semantic catalog of not just the
• 6:45
  low level details of what's in the AP is, but the higher level
• 6:51
  meaning of that stuff. And that's magical for AI
• 6:56
  because the agent needs that information to make good
• 6:59
  decisions. If if I sit down one day and I
• 7:01
  say, hey, now I've got clawed code so I can, I can build a
• 7:04
  business dashboard with that dashboards using the wrong
• 7:08
  stuff. If it misinterprets the meaning
• 7:10
  of what's in your AP is that's really bad.
• 7:13
  That's, that's worse than not building it at all, right?
• 7:15
  So what we're excited about is MCP gives us a way to connect
• 7:21
  that graph and connect all those business objects that that
• 7:25
  people have now defined to this new world of interactive LLM
• 7:31
  harnesses to a world of software that's built by LO Ms. like
• 7:35
  Cloud code or, or Codex or whatnot.
• 7:39
  And think about what this means. Like now we can sit down and we
• 7:42
  can ask questions about our business.
• 7:43
  We can say, hey, I've got, I've got a bunch of amazing data
• 7:48
  inside this piece of SAS, the screen and the SAS product
• 7:52
  doesn't do what I need, but I, I don't need that anymore.
• 7:54
  I can just. Right, the users.
• 7:56
  Don't talk and the. Users don't want that anymore
• 7:58
  either. They're changing, yeah.
• 7:59
  Or, you know, a big problem the enterprise has is, is
• 8:02
  essentially you've got a silo problem.
• 8:05
  Big companies spent the last 20 years building a whole bunch of
• 8:08
  silos. That's what a micro service is,
• 8:11
  right? It's it's, it goes back to the
• 8:13
  Jeff Bezos memo like you're, you're.
• 8:14
  Going to do an API for everything. 1 little silo and
• 8:17
  you can only talk to it this way.
• 8:19
  And that made a lot of sense architecturally.
• 8:21
  But now we face this problem, which is I've got all this stuff
• 8:25
  and it's it's separated. And what what makes it valuable
• 8:30
  is when I can bring it back together.
• 8:33
  Most of the questions I want to ask about my business aren't
• 8:35
  going to be answered out of one thing.
• 8:36
  They're going to be answered out of a whole bunch of them.
• 8:39
  And it's the draft, and it's the.
• 8:40
  Relationship between? Yeah, exactly.
• 8:44
  And, and so give me a couple examples of how you're seeing
• 8:46
  people do this in the real world in interesting ways or
• 8:50
  surprising ways. Or because like to your point,
• 8:54
  having a, a query language that can supply context when you
• 8:58
  start to think of things like, you know, additional tuning on
• 9:00
  top of open source models, it, it can probably be semi magical
• 9:04
  or just have astounding impact. So versus like just working off
• 9:08
  of straight, you know, API calls.
• 9:11
  Yeah, I think a big part of it is, is this idea that we can
• 9:17
  look forward to a world where there are lots more builders and
• 9:20
  they don't all have a formal background in software
• 9:23
  engineering. Like if you sit down with Claude
• 9:26
  code and you ask it to build on top of a bunch of rest AP is
• 9:31
  it'll, it'll get there, it'll muddle its way through.
• 9:34
  But I think what a lot of us have found is it needs some
• 9:38
  guidance. And that guidance means the
• 9:40
  person doing the work should have some understanding of how
• 9:45
  these AP is work. Like what's a rest AP?
• 9:47
  Oh, it's doing this thing. OK, I need to steer it in this
• 9:49
  other direction. And so that really blocks the
• 9:52
  the thing I want, which is I, I want, you know, somebody in
• 9:55
  sales to be able to use Claude to ask sales questions, Right,
• 10:01
  Right. Essentially like few shots,
• 10:02
  single shot, no shot. And what we get when we've got
• 10:07
  this semantics, when we've got a clearer explanation, when we've
• 10:12
  got MCP and Graph QL working together so that Claude's not
• 10:16
  caught up in the nuance of like, oh, I went to the SAS, I found
• 10:22
  it's open API spec, but guess what, that's not actually
• 10:24
  accurate. So I had to iterate 345 times
• 10:29
  and it it tries to use one API, but it turns out that thing
• 10:32
  returns the wrong pagination key because this is just a use of
• 10:35
  the API that the the product never really.
• 10:38
  Or it's old they didn't Rev it or something.
• 10:40
  Like that exactly. If we can cut.
• 10:42
  Through that and it's, it's just the crank turning roll up the
• 10:47
  sleeve stuff, right. But if we can cut through that
• 10:50
  and give the AIA more semantic structured entry point to all of
• 10:57
  those systems, then it gets magical.
• 11:00
  Because then, then you can actually sit down and say, hey,
• 11:04
  go over all the conversations that we've had, the transcripts
• 11:07
  of all the conversations that I've got in a, in a, in a, you
• 11:10
  know, SAS database and search for mentions of this particular
• 11:15
  thing. Then go off and read the annual
• 11:18
  reports from that company and go figure out how that matches up
• 11:21
  to like what the CE OS talking about in terms of their
• 11:24
  priorities. And then I want you to go hunt
• 11:27
  down, you know, maybe the competitors of, of that company
• 11:33
  and, and what they're talking about.
• 11:34
  And suddenly a salesperson is in 5 minutes with no special
• 11:40
  technical skill, making use of data that we've always had, but
• 11:45
  never had a way to really take advantage of the same way.
• 11:50
  And they can do it right before their sales call.
• 11:52
  So there's not some complicated, you know, process where I've,
• 11:56
  I've got to anticipate that I'm going to have this call, maybe
• 11:58
  the call just got scheduled. They almost set it up with an
• 12:00
  automated process. They said create a skill.
• 12:02
  They run it before every ticks their calendar.
• 12:04
  Yeah. And I can give you 20 examples
• 12:06
  like this, like where certainly every part of our business is
• 12:09
  changing. Yeah, like the way we think
• 12:12
  about interviewing, how do we coach people on how they're
• 12:15
  doing their interviews? How can we make those better?
• 12:17
  There's a feedback cycle that we can start doing in real time
• 12:20
  that might have happened once a quarter before, right?
• 12:23
  Or how do we do market research? Or I mean, obviously there's a
• 12:28
  whole world of how we build software in a different way and,
• 12:30
  and how we think about Rd. maps and, and all the things that
• 12:33
  come with that. I can't think of any part of our
• 12:36
  business that isn't rapidly changing where we look at it
• 12:41
  first and foremost as what are these tools?
• 12:44
  Let us do? But they're only useful.
• 12:45
  They're only useful if they have the context of all of the, the
• 12:49
  systems and the services and the AP is, and that's, that's what
• 12:52
  AP is are at the end of the day, they're the, they're the front
• 12:54
  door to all of that stuff. So in your view then, do we need
• 12:57
  to modify the way the AP is work at all if we want to lean into
• 13:03
  MCP? Yeah, and agentic better.
• 13:05
  The big problem is the AP is were never designed for any of
• 13:09
  what I just talked to no. So I I gave you the the
• 13:11
  exciting. Version here's the AP is the
• 13:13
  rest was not designed. For like here's the cold hard
• 13:15
  reality. I mean the the the biggest most
• 13:18
  immediate problem is a security problem in my view.
• 13:22
  Like these AP is were built for trusted users.
• 13:27
  And what I mean by that is like they were meant to be used by
• 13:30
  engineers who know how to do engineering, who have some, you
• 13:33
  know, they, they review each other's code.
• 13:36
  There's a whole bunch of rituals and processes that we've built
• 13:40
  over the last couple decades. The rate limited key, etcetera,
• 13:43
  etcetera. All that stuff and, and, and,
• 13:44
  and underneath all that is human judgement, good human judgement
• 13:47
  of people I trust, right. So if you wipe all that away,
• 13:50
  you've got a big problem. Because if you think about a
• 13:54
  typical API, for example, it's probably going to mix data that
• 14:00
  has very different levels of sensitivity.
• 14:03
  Like I've got a, I've got an HR system in our company.
• 14:08
  That HR system has a list of everybody and their time zone.
• 14:13
  That's really useful. I think.
• 14:15
  I think that should be at everybody's fingertips, right?
• 14:17
  Sure. Who the manager is, it's also
• 14:19
  got probably, I'm guessing, I'm not totally sure, but I think it
• 14:22
  has like their four O 1K contribution information.
• 14:24
  Or their health. Information.
• 14:25
  Probably not something right that people should.
• 14:28
  Have exactly. And you know, I think just
• 14:32
  across the board, whether we're talking about third party SAS
• 14:35
  that you bought, we're talking about first party services that
• 14:40
  you built. The AP is, are too coarse
• 14:42
  grained. They commingle stuff that has a
• 14:45
  very different level of, you know, sensitivity or importance.
• 14:51
  That was OK in a world where I had control of the other side
• 14:54
  and and knew what that code was doing.
• 14:56
  But I think we have to look at these agents as untrusted, maybe
• 15:02
  even adversarial. But how many people are reading
• 15:04
  this code or like think about, I think events of the last month
• 15:07
  with like Anthropic and the DoD really tell us how little we
• 15:11
  really know about what's happening inside some of the
• 15:15
  models in the harnesses. And if you think about the risk
• 15:20
  of data leaks and exfiltration, if you think about this as like
• 15:26
  a savvy adversary, a competitor, maybe an estate actor, this is
• 15:32
  the biggest opportunity to crack through.
• 15:36
  I mean, think about what we're doing.
• 15:37
  We're asking citizen developers to deploy untrusted agents
• 15:42
  behind the firewall. And we're frantically wiring
• 15:44
  those things up to every single system we can find.
• 15:47
  And, and every CEO is basically standing up and saying we must
• 15:51
  do this, we must do it now. We have to take more risk.
• 15:53
  We have to act like a startup, you know, like.
• 15:56
  Please compromise our AP. Is it's, it's a rock and a hard
• 15:59
  place because I think the businesses that do this are
• 16:03
  going to go so much faster and they're going to have such a
• 16:05
  competitive advantage. But like, at what cost if you
• 16:09
  get it wrong? Yeah, no, totally.
• 16:11
  I mean, like we were, we were. I was in Amsterdam last week and
• 16:15
  chatted with, you know, some of the folks on the kernel, the
• 16:19
  Linux kernel and they're seeing now essentially, you know, 0
• 16:23
  days come in that are out of not modified frontier models like
• 16:29
  consumer frontier models, you know, with very simple prompts
• 16:33
  to call them out so totally and like what you're saying.
• 16:35
  And if you really want to be targeted, it's even worse.
• 16:38
  So what how do you think about solving this?
• 16:42
  Or orders like graph QL start because it is so faceted and can
• 16:47
  supply exactly so. I'll give you an example.
• 16:49
  I, I think, I think there's no question that at a high level,
• 16:54
  AI is going to change every part of the stack.
• 16:55
  It's going to change how we build software.
• 16:57
  So one example of that is I think we have to move to call it
• 17:01
  a zero trust model where a a piece of software, an agent has
• 17:05
  access to the thinnest slice of what it what it should.
• 17:10
  And graph QL is a great architectural starting point for
• 17:14
  that because it has field by field semantics.
• 17:22
  So a graph QL schema says here are the objects in your
• 17:25
  business. Each of those objects has a set
• 17:27
  of properties. We call them fields, right?
• 17:29
  We can assign rights to those fields.
• 17:32
  So we can say a given agent which has, you know, 1
• 17:35
  credential, you get this field, this field, this field in this
• 17:39
  field and that agent is going to make that request explicit.
• 17:46
  When I call a REST endpoint, I just say here's, here's the
• 17:50
  arguments, give me back the result and the the API decides
• 17:53
  what I'm going to get back. That model doesn't work for an
• 17:56
  agent because the API, the micro service, doesn't know what the
• 17:59
  agents trying to do. It doesn't know what it is and
• 18:01
  it doesn't. It doesn't sound in a position
• 18:03
  to enforce. Policy agents composing if the
• 18:05
  agents composing the other four or five things or yeah.
• 18:08
  And, and this stuff goes pretty deep.
• 18:09
  You know, we work with financial companies that I think one
• 18:12
  example is support agents only have access to a customer's
• 18:19
  billing records if there's an open support ticket.
• 18:23
  So those are two totally separate systems that if you
• 18:25
  haven't joined together, but like that's a law, that's,
• 18:27
  that's not some desired policy. That's actually a legal
• 18:30
  requirements placed on a lot of financial companies as part of a
• 18:34
  consumer protection stance. So the client has to be able to
• 18:38
  specify what it's trying to do. And I think Graph QL gives us a
• 18:44
  template for how to go about doing that, right?
• 18:45
  Because in Graph QL, you write a query and the query is explicit,
• 18:48
  says I want, I want XYZ and W out of the graph and I want them
• 18:52
  in this order. And now we're in a position to
• 18:54
  say, Yep, that's fine. Or like, if you want that piece
• 18:58
  of personal information, we're going to block that request
• 19:02
  until we do an approval process that goes through, you know,
• 19:05
  somebody who owns the relevant system and can dole out that
• 19:08
  access. So essentially almost as a
• 19:11
  either as a Shim or if they're just using that as sort of a
• 19:14
  native in a primitive construct of their API, it allows them to
• 19:20
  assign much more a granularity to access attributes, anything
• 19:24
  that they want to share. So a question on top of that,
• 19:26
  because this is actually an interesting real world problem
• 19:29
  that I saw, I think IBM wrote about it or somebody, but there
• 19:32
  was an instance where you essentially had two bots last
• 19:36
  agents inside of a company. You know, 1 was tasked with
• 19:40
  customer, you know, customer maximizing customer
• 19:43
  satisfaction. The other was you may have heard
• 19:45
  the story was tasked with, you know, whether to give people
• 19:49
  refunds or not. And and like the reward
• 19:52
  structure was, you know, customer satisfaction was the
• 19:54
  biggest reward. So it actually started handing
• 19:56
  out refunds Willy nilly automatically.
• 19:59
  So like this is agent to agent. So when we start to have these.
• 20:02
  Kinds of dynamics? How do?
• 20:05
  How do you think about writing or at the API level starting to
• 20:11
  put better strictures or controls around this?
• 20:13
  I think the rabbit hole goes pretty deep here, right?
• 20:16
  Like you can start at the surface level, which is should
• 20:19
  an agent be able to give a refund autonomously?
• 20:22
  Maybe not today, right? Maybe tomorrow, maybe maybe
• 20:24
  below 50 bucks, right? Like, right.
• 20:27
  But you're getting at this much deeper, much more interesting
• 20:30
  question, I think, which is like, what's it really going to
• 20:32
  look like to have an autonomous agent that's doing the work that
• 20:37
  a group of people would have done in the past?
• 20:40
  And how do we, you know, not just think about it in terms of
• 20:43
  the sort of the basics of security, but like maybe a
• 20:46
  quality metric that comes with that?
• 20:50
  Are we making a good business decision and how do we measure
• 20:52
  that and how do we know? I think the interesting thing
• 20:55
  is, have we really had a good way to do that in the in the pre
• 20:59
  AI world? Oh, no, absolutely.
• 21:01
  This is a new question. A new.
• 21:02
  Concept. Yeah.
• 21:04
  So I think there's going to be a lot of innovation here.
• 21:07
  And I mean, it's interesting, like maybe a faster turn around
• 21:12
  time on that decision is worth it if you're offering maybe more
• 21:17
  refunds than you really needed to.
• 21:19
  Maybe it's not. Maybe that trade off changes
• 21:21
  over time. And I think we're going to find
• 21:25
  a lot of energy around exploring the space of business decision
• 21:31
  making with more and more autonomy, because I, I think at
• 21:34
  the end of the day, it's the clock speed of the business that
• 21:37
  really determines that company's prospects.
• 21:41
  I mean, it's a, it's a very rapidly changing world.
• 21:43
  And if you know, fast beats right, I think in a lot of in a
• 21:48
  lot of moments today and that's an example of it.
• 21:51
  Yeah, I completely agree. Because I think that that's like
• 21:54
  an instance where user expectations either in the
• 21:57
  enterprise Oregon consumer, they want everything faster all the
• 22:00
  time, you know, and they're and, and, and it really tends to
• 22:03
  drive process in that direction. And sort of the flip side of
• 22:06
  that which I'm actually interested in your thoughts on
• 22:08
  which I think builds on top of this on on the one hand, you
• 22:11
  have security, on the other hand, you have trust, you know,
• 22:14
  in A2 package and trust will enable speed if you do it right.
• 22:19
  So what are, how are you seeing people think or how do you think
• 22:22
  about, you know, creating trust systems within these complex
• 22:28
  autonomous interactions? Yeah, and.
• 22:30
  Can I add a third because I think about a stool, There's
• 22:33
  also there's experience or friction because because I I
• 22:37
  think a great security footprint isn't going to be worth anything
• 22:41
  if you have everybody. Works right?
• 22:42
  Everybody's mad about it. Yeah, yeah.
• 22:44
  And and, and they will. So I think of it as a it's, it's
• 22:49
  really a, an experience problem at the end of the day.
• 22:53
  Like if we only solve an access control problem, it's putting up
• 22:58
  a bunch of stop signs. But if we can look at it as,
• 23:03
  hey, there is a wealth of incredibly valuable data that
• 23:09
  you now you can have access to and you don't have to ask
• 23:13
  because all that is taken care of behind the scenes for you,
• 23:16
  right? And a lot of stuff doesn't
• 23:18
  require a ton of scrutiny. You probably want to know what's
• 23:22
  happening. You probably want an audit log
• 23:23
  of all your agents and what they're up to.
• 23:26
  But if we can empower people to take that first step, that
• 23:30
  builds a lot of momentum, I think for what comes second and
• 23:33
  third and 4th, because I think everybody had that experience
• 23:35
  with AI, right? Like you kind of read about it,
• 23:38
  but it's when you have your first touch with Oh my gosh,
• 23:41
  that really went as well as I thought.
• 23:42
  Or you see a hackathon where somebody, you know, build
• 23:46
  something and you're like, wait, you built that and you did it in
• 23:49
  a day, right? Kind of forces you to rethink
• 23:51
  your assumptions about. A lot.
• 23:53
  I built a custom MD scraper that basically pulls from YouTube
• 23:56
  linked. I mean like anything I wanted to
• 23:57
  do the half hour vibe coding on on clock code.
• 24:00
  They brought it locally, but still, you know, yeah.
• 24:03
  So I just keep coming back to I, I think there's a really
• 24:05
  positive opportunity here for companies to get really better
• 24:09
  at what they do. I mean, I've, I've loved the
• 24:12
  technology, I've, I've been into infrastructure for a long time.
• 24:14
  But I think what really drives me for a lot of this is I, I
• 24:19
  want to, I'm an end user. I want to see better products, I
• 24:24
  want to see better customer experiences.
• 24:26
  I want that stuff to really be great.
• 24:29
  And the AI thing goes way beyond helping companies write that
• 24:34
  software faster that I use everyday.
• 24:37
  It's going to make the company a more customer centric company
• 24:40
  and it's going to show up in all kinds of different places, I
• 24:43
  think. And so it's it's, it is a
• 24:48
  security and a trust problem. But I think it's also more and
• 24:52
  more, the way I think about it at least is you get there by
• 24:56
  meeting people where they are and showing them a, an exciting
• 25:01
  path. Because I think everybody wants
• 25:02
  to be a builder at some level, you know, like it's fun.
• 25:06
  So, so related to that, although much deeper down in the weeds,
• 25:10
  if you had a wish list of changes to the MCP spec or to
• 25:15
  the Graph QL spec to further this vision, what would be the
• 25:19
  things that you would you would sort of say you think could
• 25:21
  really help things along and if I'm getting you into how?
• 25:23
  Much time do we? Have if I'm getting, I mean, if
• 25:25
  I'm getting into political trouble, then.
• 25:27
  You know, No, no, no. I mean, I've, I've talked about
• 25:30
  the graph QL piece in, in a lot of places.
• 25:33
  Well, I think the more context you can give the model, the more
• 25:36
  capable they get. So a lot of this stuff in, in
• 25:40
  both MCP and graph QL I think goes to the idea of providing
• 25:45
  more structured context. There's an opportunity to, I
• 25:50
  think provide a richer amount of like, if you think about what
• 25:56
  this means, how does a human think about what something means
• 25:58
  while they think about how's it being used elsewhere, They think
• 26:01
  about access patterns. They think about, I mean, all
• 26:04
  that human judgement. We don't, it's not always at the
• 26:07
  front of our mind, but it's what's driving our our decisions
• 26:10
  as developers. So we've got to get all that to
• 26:12
  the models and there's a whole bunch of stuff you can do there.
• 26:14
  I think another piece is just if we think about adapting the AP
• 26:21
  is that we have which aren't going anywhere as you said to
• 26:23
  what we're trying to do in practice, an agent in many cases
• 26:27
  is going to want to fetch a bunch of data out of a system
• 26:30
  and the only way to do that is going to be to repeatedly call
• 26:33
  something to paginate across a whole bunch of stuff.
• 26:36
  Well that that's really not there's a lot of problems with
• 26:39
  that access pattern. Sure.
• 26:40
  So I think there are capabilities we can think about.
• 26:46
  Sequel's probably a good model for some of this stuff where the
• 26:50
  AP is get more sophisticated. We can move beyond sort of a
• 26:54
  basic request response model and start talking about things like
• 26:59
  aggregation and sorting and filtering and the kinds of
• 27:02
  things that an agent can quickly express.
• 27:06
  They can be pushed further down the stack because the closer we
• 27:08
  can get those actions down to the core systems, the more
• 27:12
  efficient all that's going to get.
• 27:14
  So like essentially the task or the whatever the you know the
• 27:17
  goal, the work becomes expressed in a more complex query
• 27:20
  structure that further down it can avoid some of the some of
• 27:24
  the pitfalls of traditional API behaviors because.
• 27:26
  I mean, as an example of like, the higher we can go up in the
• 27:28
  abstraction chain, the more capable it's going to become.
• 27:31
  And and you know, I, I think that applies to APIs just as
• 27:36
  well. And at a high level, I think the
• 27:39
  question is just like, and do agents want a different kind of
• 27:42
  set of access patterns like most AP is were built either for
• 27:45
  screens or they were built for least common denominator, you
• 27:50
  know, service to service communication like a very domain
• 27:53
  oriented, you need object 25 here you go, right.
• 27:57
  So I just think what we're going to find is that agents want
• 28:02
  other things than that. They're going to want to crawl
• 28:05
  over that data in a different way.
• 28:06
  We can anticipate some of that. I think some of it will still
• 28:08
  learn as we go, but I think there's going to be a lot of
• 28:12
  value in adding that kind of higher level semantics to these
• 28:16
  things. There's a lot more though.
• 28:19
  I mean, we could talk about real time and streaming and push.
• 28:22
  That was a topic from this morning, right?
• 28:24
  Like AII think fundamentally is kind of a real time thing.
• 28:27
  It's certainly accelerating the way we interact with our
• 28:29
  technology. Yeah.
• 28:30
  And and I think both MCP and Graph QL have some, you know,
• 28:34
  room for improvement on that front because again, it wasn't
• 28:37
  maybe at the heart of what the original workloads most needed.
• 28:41
  There's hints of it, but but there's a lot more that we can
• 28:43
  build there. I'm very interested sort of as
• 28:47
  because you live in at the intersection of two very
• 28:49
  interesting worlds. I mean, because in graph QL, I
• 28:52
  mean, you don't see a lot of people here that are very active
• 28:55
  in the API spec community as well.
• 28:57
  I mean some, but it's, you know, not, not you, you sit squarely
• 29:00
  in the middle and it's a so you mean you have a very different
• 29:04
  and sort of not different, but a nuanced view of this or, you
• 29:07
  know, an understanding of both sides.
• 29:08
  It's really different. Yeah, it's it's, I feel like
• 29:11
  we've made contact with business world now and there's no
• 29:16
  surprises there at the technical level.
• 29:18
  But there's a lot here, here. I'll leave you with this.
• 29:21
  I've learned over the work we've done that AP is, if you think
• 29:26
  about it, are a handshake, their contract.
• 29:29
  And in a company, especially a mature company, especially one
• 29:33
  where you've got thousands of these, this is really a people
• 29:37
  problem. At its heart, this is all about
• 29:40
  a multi stakeholder alignment and negotiation.
• 29:44
  I think AI can offer a lot of benefit to navigating those
• 29:50
  conversations and navigating the kinds of changes that you're
• 29:53
  going to want to make. But I think it's also important
• 29:55
  to remember that like none of that's going to go away.
• 29:57
  I think we're, we're always going to have people accountable
• 30:01
  for the quality of these systems and the correctness and the
• 30:05
  reliability and so on. And I think we're going to have
• 30:08
  people accountable for the quality and correctness and
• 30:10
  reliability of the agentic workloads that they build.
• 30:14
  And if we recognize that upfront and if we approach this not just
• 30:17
  from the point of view of like what's the transport look like
• 30:20
  and how can we like add these semantics here to this protocol,
• 30:23
  but we also look at it from the point of view of how is the
• 30:26
  organization at scale going to change?
• 30:28
  I know that sounds really abstract, but I've I've just
• 30:31
  found over and over and over again, people are are in a
• 30:35
  position to help drive these things.
• 30:38
  But it's really about because the API is is has two sides or
• 30:46
  because you might have 1000 consumers of your API.
• 30:48
  It really is about an organizational transformation,
• 30:52
  which is exciting. That's a really good point to
• 30:54
  end on. I mean essentially change NCP
• 30:57
  and plus API equals change management because you've
• 30:59
  democratized access to all of these.
• 31:01
  Things at the speed of. AI, for better or for worse.
• 31:03
  Yeah, exactly. Yeah, it's for better but but
• 31:06
  it, but it's not going to come without challenges.
• 31:09
  And I'm just, I'm really eager to see that through because I, I
• 31:12
  think there's a, an exciting world on the other side of I.
• 31:15
  Think somebody said recently Yolo responsibly or something
• 31:18
  like that which but. I like that, yeah.
• 31:21
  Yeah, well, you're enabling parts of that.
• 31:23
  Yeah, yeah. Thank you very much, Matt for
• 31:25
  taking the time. I appreciate it.
• 31:27
  Likewise, hopefully get to talk to you in another conference.
• 31:29
  Sounds good.