The rise of the agentic economy on the shoulders of MCP

Talk transcript

1. 0:00

   [on hold music] So let me start with a question.

2. 0:16

   How does intelligence emerges in, uh, biological systems, right? Well, it's through neu-neurons, right? Well, when neurons are born, they are just like individual cells. But like over time, they grow their axons and dendrites and establish connections with other cells or other neurons and actually learn how to communicate in order to pursue their own interests, basically, like to

3. 0:39

   get nutrients and so on. And over time, they learn how to, how to communicate with each other and with other cells to get nutrients and basically thrive, right? And this collective behavior, if you like sc-zoom out and look at like really large number of them, uh, it's something we call intelligence, right?

4. 0:54

   So it's like emergent behavior of smaller individual units that pursue their own interest. So how does intelligence emerge, uh, emerge in the markets, right? People always talk about markets like, well, market thinks that, market, uh, reacted to this, and so on.

5. 1:11

   And in some way, uh, markets are more intelligent, uh, than like individual like participant of the market, right? And it's-- there are mutual, uh, interaction of these individual members of the market, basically, who pursue their own interest and communicate and establish new interactions with others, uh, where some, some sort of like collective intelligence, which is like bigger

6. 1:33

   than the sum of different parts emerges, right? So how does intelligence emerges, uh, emerge in companies? Well, this one is provocative, through Slack, right? [laughs] Where people interact and pursue their own interest, uh, in the company and over-- like all together, the company, well, sometimes becomes more intelligent than the individual employees of the company.

7. 1:55

   And, uh, so this leads to my final question. So how does or how will the general intelligence emerge in computing systems, right? And there is a lot of talk about AGI and, uh, like, you know, like ever larger models, uh, exhibiting like, uh, super intelligent behavior.

8. 2:11

   But in my opinion, the like general intelligence will actually emerge through interaction of multiple entities, you can call them agents, basically like multiple models, uh, pursuing their own goals, interacting with each other, and, uh, all together exhibiting something which we can call general intelligence.

9. 2:30

   And thanks to, uh, MCP, we finally have this, uh, missing part that allows the, the agents to communicate with each other and really like create a fabric or agentic mesh where they can talk together.

10. 2:43

    So, uh, hello everyone, my name is, uh, Jan Curn. I'm the founder of Apify, and, uh, I'm gonna talk about the rise of the agentic economy on the shoulders of MCP.

11. 2:52

    Basically, economy where agents, uh, can, you know, find counterparts, uh, to interact with and purchase services from other, uh, uh, from businesses or tools or other agents, right? So like B2A and B2B, uh, sorry, and A2A.

12. 3:09

    All right. So before I start, um, let me just introduce quickly Apify. Apify is a, is a marketplace of five thousand tools called actors. And, uh, historically, we come from the web scraping industry, right?

13. 3:22

    So most of these actors are data extraction tools that allow you, you know, to get data from social media, from search engines, uh, data for AI, for building RAG pipelines, you know, uh, data from web, uh, for lead generation, and so on.

14. 3:36

    But also there are other tools like data processing tools and so on. So altogether, there's about five thousand of them, and some of them are built by Apify, some are built by our community of creators, uh, who actually make money on it, right?

15. 3:48

    So it's like a marketplace of software creators, if you will, right? So actors are self-contained piece of-- pieces of software based on Docker with well-defined input and output, right?

16. 4:00

    And basically, they represent a new way how to ship software and publish it, you know, and, uh, and integrate to, to, you know, other systems, right? So for example, Google Map Scraper, it's quite popular actor, uh, from our store.

17. 4:12

    Uh, it can extract data from Google Maps, right? Uh, more data than, than the Google Places API provides, right? Uh, well, there is like creator of the actor description, you know, uh, different stat and so on, something you would expect from normal mar-marketplace.

18. 4:25

    And actually, thanks to the way actors are built, it's actually super easy to integrate actors from other systems, right? So for example, we have SDKs for TypeScript, for Python, uh, for OpenAPI, uh, for C-- um, um, for CLI, it's like we can call it from terminal.

19. 4:43

    And it's only because they are well-defined units of software with input and output, right?

20. 4:48

    Uh, also we have integrations with, uh, workflow automation tools like Make, Zapier, you know, Clay, and many others. So to make it really easy to call actors from these systems, right?

21. 4:59

    But obviously now, uh, we also have MCP integration, which makes it possible to call actors from AI agents or, you know, AI workflows. And the way it works actually is, uh, the agent just needs a API key or, you know, workflow on an a-an account on Apify, and then through our MCP server, basically it can interact or

22. 5:24

    call any of those five thousand actors on our marketplace, right? And actually, this only became possible thanks to, uh, I would say the killer feature of, uh, MCP, which is the tool discovery, right?

23. 5:36

    Actually, um, not many clients support this yet, uh, but, uh, just yesterday, I saw that V-VS Code added support for it. Uh, and actually just like two days ago, Claude, uh, for Desktop added support for tool discovery.

24. 5:49

    And basically how it works is that, um, the client connects to the MCP server and dynamically discovers tools to use and to interact with based on the, based on the, the, the workflow, right?

25. 6:03

    And let's say we have like five thousand tools on our, our store, and there is simply no way we could publish all these tools through OpenAPI because, you know, the, the context would be just too large.

26. 6:12

    And like the more tools you have, the, you know- ... riskier the result is, right? So we really want, like, provide the tools only, like, uh, as needed. And that is only possible t- through tool discovery, which I think is really the main thing that will actually make MCP really, uh,

27. 6:31

    the huge differentiator from, from OpenAPI, for example, right? So MCP actually quickly became a standard for agentic interaction. This is, uh, Google Trends data showing that MCP is, is basically dominating the space compared to OpenAPI or A2A from, from Google, right?

28. 6:47

    And actually, I think MCP already became a standard for agentic interaction. And it became so popular that w- currently there are, like, you know, many different like, uh, registries of MCP servers that even guys from Master, our friends, created like registry of MCP server registries, right?

29. 7:06

    Just to [laughs] make the sense of it, right? And obviously, Anthropic is also working on their own, uh, registry. And, uh, I think G- Google's A2A, they have like a DNS-based protocol with like well-known .agentjson way to, you know, publish the, the services on t- through DNS.

30. 7:24

    So basically, there is like, you know, so many different servers you can now use from the agents, right? So does it mean with like... So many tools now support MCP, so does it mean like the agents can discover and access any of them on their own, right?

31. 7:41

    Well, not really, because to use those services, your agents still need, uh, to have like API tokens to those services, right? So even if... Let's say if you use, uh, Zapier M- MCP that provides access to like 5,000 apps they have in their marketplace, you still need to connect those individual apps to your services, right?

32. 8:00

    You know, to like GitHub or Slack or, you know, whatever. So Zapier on their- on its own is not able to provide access to this, uh, th- third-party services.

33. 8:09

    You still need to, as a user, to facilitate it. So that, that actually means that, uh, the agents are not able to, like, find counterparts, uh, or like, uh, other agents or other tools to interact with on their own.

34. 8:23

    They are still depending on the human, uh, developer who actually build the system, right? Who kind of like give those, those, those agents access to different tools, right? And if those agents are, you know, to replace all the people and all the jobs, right, they s- need to be able to, uh, find services to interact with.

35. 8:41

    They can't just like, you know, do, do that. Like, it's like s- basic, basic thing that like, uh, a- anyone of us can do, right? Like to find service and purchase it, right?

36. 8:51

    So I argue that like unless the, the, the agents are able to do that, uh,

37. 8:57

    we will not be able to reach, you know, some higher level of, of intelligence of these agentic systems and behaviors, basically, uh, if the agents cannot purchase services, right?

38. 9:07

    So how can we solve this problem, right? So first, like, sort of like naive approach would be let the agents subscribes, uh, themselves to the target services, right? So basically, in a way, like agents could have like email, maybe a credit card.

39. 9:21

    They could like fill, you know, the subscription flow, maybe solve the CAPTCHA, you know, create an account, and so on. But you see, it's, it's not very practical, right?

40. 9:28

    I mean, it's, you know... Well, they might need to also have the phone number, and so on. And quite often the services actually need to have like real person behind the account, right?

41. 9:38

    So basically, this, this wouldn't really work, right? Uh, so second solution, uh, is central identity and payments provider. There are like couple of companies pursuing now that, like, there would be like a central authority where you can charge money, and then the agents can use that, you know, to buy services and, and provide them with their identity,

42. 9:56

    right? For example, Vertify. Coinbase is now pushing their X402 standard. Uh, I think Stripe is working on this and MasterCard and Visa too, right? So I think this is gonna, this is gonna happen eventually, but running or launching new payment system is extremely complicated, right?

43. 10:12

    Because you are facing like this chicken and egg problem of marketplaces, right? I think PayPal had to, uh, pay like $100 million per month just to buy the market.

44. 10:20

    Uh, and like launching credit cards in the, in the '70s was like incredible challenge basically because nobody was accepting those cards, so why would people use them, and so on, right?

45. 10:28

    So I think this will happen, but it will be long process basically to establish this, right? So let me offer the third approach, and it's like through a centralized, uh, marketplace of MCP services, like Apify Store basically, where you just need one API token or one authentication, one account to, to get access to all the other services.

46. 10:47

    And basically, it works the way that the developers who publish these tools, these actors, actually they provide their credit card and their account to the third-party service and basically publish it, add monetization to it, like, uh, like how much does it cost to call this service?

47. 11:02

    And then they are basically the owner of the service, and now they publish it on our marketplace, and suddenly it becomes available to the whole ecosystem of tools. And this way, actually, we can scale it rapidly and actually even without the target services knowing, right?

48. 11:15

    So basically, this way, the actor can co- run the code itself or wrap an external API or just publish an external MCP server. Because the MCP servers, they can be actually nested.

49. 11:25

    You can have like one parent server that provides actions or tools of the, like, nested MCP servers, right? So that's another, like, cool feature of, of MCP. You can really build this sort of ecosystem, you know, if you can facilitate the payments and monetization, right?

50. 11:39

    So actors charges the user, and then it, its developer gets the money and pays for the external service. And anyone can publish su- such an actor even without the target service knowing, right?

51. 11:51

    So time for demo. It's not live demo because the internet, uh, super flaky here. So what you can see here is Claude for desktop, uh, that has access to Apify MCP server.

52. 12:03

    Uh, there is like 18 tools available now. And I'm asking like, "What is the venue of AI Engineer World's Fair in San Francisco? If possible, use Apify Actors." So it...

53. 12:11

    You can see it searches the actors, uh, for a tool that can answer this question. It will find a tool or actor called RAG Web Browser. And so it's call- it's, uh, it's like a Google search with, uh, you know, fetch data.

54. 12:25

    So basically it, it asks the query, like, uh, what is the venue and so on, and then it pr- parses the, the resulting page. So we can see it found, like, uh, SF Marriott Marquis.

55. 12:35

    Uh, that seems all correct. Right? So now let's use an actor for scraping Twitter.

56. 12:43

    So, uh, this actor is not available in the context, so, so the agent doesn't know how to use it. So it will... It searches actors on our store and finds an actor that can, uh, scrape Twitter, right?

57. 12:54

    So it, it calls, it calls add actor, which is like a tool that adds new tool to the context. Uh, actually, Claude is very verbose, uh, describing a lot of things about it.

58. 13:04

    And actually, there is, like, small box still in Claude Desktop that you need to, like, disable and enable a tool so that the, the tool, uh, list refreshes, and then the tools become available.

59. 13:13

    I'm sure it's gonna be fixed in the next release. And now let's use that actor to, uh, get last tweet of AI engineer conference. All right? So it calls the actor on Apify.

60. 13:23

    Uh, it knows the, the Twitter handle, probably from, from the, from the website. And now you can see that, uh, it found result and the last tweet from, uh, this morning was, uh, something about workshops.

61. 13:36

    That seems about right. So now what? So we, we have, we have seen how we can use existing tools in, in, in our store, but, like, let's say, uh, uh, one of our competitors, a company called Browserbase.

62. 13:52

    Hey, Paul, if you're here. Uh, they r- certainly haven't published, you know, uh, an actor in our store, but we did. So we created an account on Browserbase, added our API token there, and published, like, basically their MCP server on our store without actually them even knowing.

63. 14:09

    And now anybody can actually use Browserbase MCP through Apify's ecosystem, right? Even without them having to do anything or knowing about it, right? So now let's use Browserbase to fill in the email subscription form on the AI engineer website.

64. 14:25

    Fill email, uh, [REDACTED:email_address]. And now let's see what happens, right? And actually we'll see that, uh, that the, that the agent will actually call Browserbase MCP through, you know, an actor published, published, you know, by us on our, on Apify Store, and perform the actions on the web, right?

65. 14:43

    And actually, this way we can easily, like, uh, bring a lot of, lot of existing MCP servers to our store and, you know, expand the ecosystem rapidly without, you know, having to ask for, for, you know, cooperation of those third parties, right?

66. 14:58

    So that's actually what we're doing now. Uh, we want to scale this marketplace rapidly. And now, okay, so now it's evaluating, you know, the screenshots, looking for the field, and so on, you know.

67. 15:09

    And eventually it will manage to, uh, fill the form and, and basically succeed in the task, right? I can maybe skip this, uh, to save time. It takes some, some, some time to basically, uh, for the agent to, to find a form and so on.

68. 15:26

    But, uh, yeah. It succeeded. It completed the email sub- sub- subscription, and

69. 15:34

    this way you basically see that, uh, you can plug

70. 15:40

    our, our ecosystem of actors into, into, uh, any AI a- agents that, that actually support tool discovery, right?

71. 15:47

    All right. And... So this means now anyone can publish tools or, you know, agents on Apify Store and monetize them and immediately get access, you know, to all the AI clients that already, like, integrated with Apify and all the ecosystem of tools, right?

72. 16:08

    And actually people can ma- make money on it. Like, just last month we paid, uh, more than a quarter million dollars to our creators, and actually the, the, the...

73. 16:16

    this number is, is growing rapidly. You know, overall the actors generate more than one and a half million dollars per month now. Uh, we have, like, uh, one million monthly visitors to the whole ecosystem, and now we're really in the process of, like, scaling this ec- ecosystem.

74. 16:29

    So, um, if you're looking for ways to monetize your tools or agents, you know, just, uh, talk to us and publ- or pub- Apify Actor, Apify Store and get access to this ecosystem of developers and this visibility.

75. 16:44

    And there are some open questions obviously, uh, that remain. So will this autonomous tool discovery provide real value? I mean, like, everybody who builds agentic systems knows that, you know, like, making sure that the system works as expected is tricky, right?

76. 17:02

    Even if it's fixed. So if we add this, like, you know, variables that, like, well, uh, if the agents can discover new tools, uh, will it actually work? Well, currently it might, it might be a bit flaky, right?

77. 17:14

    I think we're still, still, still, still... We, we're still fairly early. But as the models get better,

78. 17:20

    I think, uh, even with the discovery, suddenly the, the, the agents will be, will be able to provide, you know, valuable and w- reliable result basically, right? So it just r- this remains to be seen, but I'm optimistic that, like, as the, as the LLMs will get better, we'll actually get there, that the tool discovery will actually

79. 17:39

    provide real value. Well, there's a big question of, like, how can agents trust tools or other tools... Or, sor- sorry, or each other, right? We know it, like, you only interact with people you trust.

80. 17:49

    So how can agents do that? We'll see. And can autonomous agent interaction enable AGI? Well, we'll see. Thank you very much for your attention, and, uh, feel free to try it, uh, mcp.apify.com. [outro music]