Agents Need Receipts, Not More Tool Calls - Armanas Povilionis, Alithea Bio

AI Engineer · 19 min · 165 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:00Hello everybody.
2. 00:02I'm Armandus.
3. 00:03I work at Alifeia Bio where we focus on life sciences topics.
4. 00:07More specifically, we work on immuno-oncology and immunopeptidomics.
5. 00:13While we work on science topics, we use ourselves and provide our clients agentic automation or agent AI solutions.
6. 00:23However, we found a certain gap which we would like to discuss today and we have a solution which we're working on which we would like to share.
7. 00:34It's called Froglet.
8. 00:37But to start with, I would like to ask what do you think is most valuable work that agentic automation can do?
9. 00:49My guess, uh scientific research automation would be quite high up that list.
10. 00:56However, from my experience, adding more tools to agents will not suffice.
11. 01:03Because science and especially lives in in specially in life sciences is inherently collaborative process.
12. 01:14And in order to enable agents to collaborate, we need a way to have a verifiable chain of receipts.
13. 01:24What do I mean?
14. 01:25We need a solution which can provide these receipts proving every step, ensuring that every result can be trusted, and enabling repeatability as well as collaboration at scale.
15. 01:41Having these pieces in mind, let's go to the next step.
16. 01:45So, stepping back, um we imagine agents as cooks in the kitchen.
17. 01:54Giving them more tools or better tools, it improves kitchen's efficiency.
18. 02:03Better knives, more pans, more ovens definitely boost the speed and quality.
19. 02:10But it only enhances the local work.
20. 02:14On the other hand, scientific work is not cooking alone in your own kitchen.
21. 02:21It is closer to running a Michelin-star restaurant.
22. 02:26To overcome this the the outcome depends on providers and quality of your produce, high-level services, as well as an an ability to consistently deliver the same quality dish again and again and again.
23. 02:45You cannot bring everything into one kitchen.
24. 02:50The challenge isn't local.
25. 02:51It's not local tools.
26. 02:53The challenge is aligning the entire supply chain that it's repeatable and consistent.
27. 03:02Today, we have already agents with plenty of tools and increasing in volume, number, and quality.
28. 03:12And we also have on the hand data and specialized analytics algorithms and compute which is distributed across different organizations and siloed.
29. 03:26So, our vision is that whenever AI agent workflow automation matures, organizations will give to the agents not only tools, they will give them budgets.
30. 03:43We're already kind of doing this in a primitive form by allocating tokens for token budgets for a specific task.
31. 03:54But, what I'm talking is a bit broader.
32. 03:57It's allowing agents to manage their own budgets to achieve their own goals.
33. 04:01Whether it's to discover services, request data, negotiate execution terms, or pay for work across organizational boundaries.
34. 04:13At that point, an agent is no longer just a cook with a better knife.
35. 04:19It starts to act like a chef.
36. 04:24Uh like a executive chef, uh finding suppliers, ordering ingredients, coordinating work at at the kitchen, and finally keeping the record of what has happened.
37. 04:38That is why we're building Froglet.
38. 04:41An open-source protocol for agents to discover, transact with, and receive verifiable receipt from external data uh data and service providers.
39. 04:56For more detailed information, obviously, go to our page froglet.dev, uh where you will find um general descriptions on how it works, what it integrates with.
40. 05:09It Froglet is designed to be in between the moving parts.
41. 05:14It's not designed to replace, it's designed to integrate.
42. 05:17Integrates with different payment rails, it integrates with different agent harnesses, execution environments, and even transport protocols.
43. 05:27And it doesn't require that every node in the network uses the same software stack.
44. 05:33That's the beauty.
45. 05:33We are not forcing everyone to use the same to to work the same way.
46. 05:38We're just asking everybody to use the same interface.
47. 05:43In the web page, you will find how to run froglet with just how to install it locally just with one command.
48. 05:53Or even how to run it remotely just with one prompt.
49. 06:00So in essence, what we are saying is closed source and the collaboration often turns into a bespoke project, enterprise project that can take years and cost millions before any reusable
50. 06:14workflow exists.
51. 06:18On the other hand, froglet's mission is to make transaction layer much lighter.
52. 06:24Once an organization decides that the data or resource or service is deemed shareable and it's exposed on froglet, an agent can discover it, understand the terms, request the work, and receive
53. 06:41a verifiable receipt.
54. 06:46That setup costs few thousand tokens and takes minutes.
55. 06:55Now, if I may, let me open our website for quick second.
56. 07:03So if you land our if you go and land our homepage, you will see a couple of things.
57. 07:10First, a lot of helper information which allows you or your agent to read through and understand how it works.
58. 07:18And most of the webpage is focused on agent use.
59. 07:23But also we have some helpful documentation for you.
60. 07:27Let me go through a couple of steps in the demo.
61. 07:32So first of all, we define that froglet node by itself is not different implementations at each site.
62. 07:41It is actually exactly the same node.
63. 07:44It just plays different roles, whether it's a provider, requester, or marketplace.
64. 07:50Marketplace itself is just a Froglet node, which is providing certain services.
65. 07:59And what does it help to do?
66. 08:00It It helps to find, to trust, to pay, and to prove that the work has happened.
67. 08:09That is it.
68. 08:10We're not trying to replace any of functionalities of other tools or protocols.
69. 08:18Whenever you are generating a or creating a new node, it generates a key pair for identity and signing the artifacts.
70. 08:27And during the execution, everything is being signed with that signature.
71. 08:33Meaning, everything is being signed in a chain, and the chain is only valid if all data points are not tampered with.
72. 08:45And in in terms of discovery, request uh providers are publishing what services they uh they uh they're providing to the marketplace using the same protocol.
73. 08:58And requester requests a known marketplace just to provide lists of services that are available on that marketplace.
74. 09:08From that that moment on, once the requester identifies the service that they want to work with, the communication is direct.
75. 09:17There is no uh third party.
76. 09:20And everything from quote to deal to execution and receipt happens in one interaction.
77. 09:30Let me jump forward a bit.
78. 09:34So So far, what I showed you is just a node talking to a node.
79. 09:39Where agent comes in uh are the plugins that we placed inside the Froglet core.
80. 09:47It integrates with different execution environments and different harnesses, whether it's Open Claw, Nemo Claw, it's acts as an MCP server or plugin, and it's primarily designed that it is
81. 10:04used by an an agent.
82. 10:08The primary interface for humans should be an LLM.
83. 10:12An LLM should drive usage of Froglet, whether you are a provider or a requester.
84. 10:38So now maybe let's jump to a terminal and I will show you a couple of things.
85. 10:43First of all, I will show you how to how you can use Froglet remotely, meaning that there is a Froglet node running in a remote server where you can access and get a 15 minutes
86. 10:55trial identity which you can play with.
87. 10:59Second, how to run a Froglet locally on my machine here on a Docker.
88. 11:04And third, how I will ask my Claude to interact with local installation and actually configure a service and try it out.
89. 11:14So first of all, what we what I would like to do is to run a command to check whether Froglet try.froglet.com.dev actually exists and it works.
90. 11:27That's fine.
91. 11:28So now what I can do is to create myself a token.
92. 11:35And this this creates a temporary token for 15 minutes.
93. 11:40And you can see it here created.
94. 11:47And then what I can do also in exactly same manner, I can have access to the specific provider token.
95. 11:56And this is a specific provider which adds two numbers.
96. 12:01It's very simple example service provider.
97. 12:05As you see that the tokens are different.
98. 12:08That means that froglet nodes are different.
99. 12:12Now, in order to execute this service, what I need, I need to create a payload.
100. 12:22So, I created payload which includes many things including schema, what kind of service I'm targeting, also which provider, and finally what input I'm providing it to.
101. 12:37So, I'm providing it seven and [clears throat] five, 7 + 5. Now, what I need is to provide this payload to my remote remote froglet.
102. 12:54And this executes, I need to get a deal ID.
103. 12:59And this is where I have a receipt of actually deal happening.
104. 13:04Now, deal has an ID already, but it might have happened
105. 13:11[snorts]
106. 13:11that it takes some time to so execute.
107. 13:13So, I take a for command for loop to actually wait for it, but it depends on the load on that temporary small node that uh that runs on on the cloud.
108. 13:25And here we have an answer.
109. 13:26It's 12.
110. 13:27Fantastic, we have a correct answer.
111. 13:30Um, so we have a entire loop finalized.
112. 13:34Now, in order to run it locally, um what I need is to install it here.
113. 13:42So, first of all, I will clear everything.
114. 13:46As you see, the uh the the directory is empty, and what I do is I what I run is HBS Froglet Dev Agent, and what it will do, it will download and install agent
115. 14:01on a Docker.
116. 14:04Let's wait a second.
117. 14:05It's actually already there, so it just confirms that it doesn't need to download anything and it's working and it's running.
118. 14:15And now I can see that uh there are two Froglets running, um but if I look at at the actually not IP addresses and ports, but I actually look at the node ID,
119. 14:29I see that this is one and the same ID.
120. 14:32And that means that each each Froglet can assume different roles, and at the same time can assume multiple roles.
121. 14:40And we have here two roles as a provider and consumer at the same time.
122. 14:45And just to prove that I'm running it on a Docker, here it is, running on the same ports, same IP addresses, my local ins installations.
123. 14:57So, now Claude, if I open my Claude, first of all, what I need to do uh is to see where Froglet MCP is enabled, and it is.
124. 15:11That's fantastic.
125. 15:13And what I will ask it to do is, first of all, I will change the permissions that it doesn't bug me.
126. 15:24And what I will ask it is use from from that um CP only.
127. 15:35Do not install anything.
128. 15:42And um let's be very specific.
129. 15:48Let's see if it understands constraints.
130. 15:52And I have here prepared a couple of steps.
131. 15:56I won't do it them one by one.
132. 16:00I'll actually ask it to do it all at once.
133. 16:07In steps and provide the specific output.
134. 16:15Great, it even goes ahead and um uh checks what is running.
135. 16:21And what I'm asking it to do is first of all to see what was the status is life, which already did above.
136. 16:31Then if provider and runtime is healthy, then I want it to publish an artifact from a template, which is exactly the same template add two numbers.
137. 16:42[snorts]
138. 16:42Then uh what I want it that it actually finds and calls that service and invokes it locally.
139. 16:51Uh asking it to add again five and seven.
140. 16:56And it can wait a bit because it's an asking for a call.
141. 17:02And then show the output what it got.
142. 17:05So if if we go if we go down here, here is what we see that it actually executes all commands.
143. 17:19And what we see that is actually um done exactly as we asked.
144. 17:24We It has a result some 12.
145. 17:27And it It can create anything that you want.
146. 17:31It can connect to You can ask load to connect to your database and provide row by row access for for uh for a payment or you can have a just your GPU running in the background and
147. 17:46welcome the requests for computa- any kind of computation.
148. 17:51In essence, it's a very simple way how to use it, but it packages a lot of um underlying protocols and underlying tools where now you're not shoving everything into
149. 18:04a uh LLM context.
150. 18:06It just has a services that it needs to interact.
151. 18:10And underneath there is a receipt, there is a payment, there is a execution environment.
152. 18:17And uh there is a negotiation and marketplace uh hidden and it's uh doesn't stuff your LLM with context.
153. 18:26So, going back to the slides.
154. 18:31Some reason it scrolled down.
155. 18:34Going back to the slides.
156. 18:37Um So, what we what we have uh on our hands.
157. 18:42Uh we have an open protocol called froglet which um has multiple things.
158. 18:47First of all, uh it uh provides a way uh to to find resources outside the organizational boundaries.
159. 18:59Then it allows your agents to actually execute uh remote uh commands.
160. 19:06And third, uh what it allows, it allows to have a verifiable receipt of receipts of what has happened.
161. 19:14And I think that actually enables a collaborative science.
162. 19:19That enables a progress of a perpetual science automation.
163. 19:26So, please join us either via Froglet or Althea Bio.
164. 19:31And please join us on GitHub.
165. 19:33And thank you for listening.