A2A Protocol (Agent2Agent) Explained: How AI Agents Collaborate — Transcript

Full transcript

1. 0:00I may have brought up the topic of AI agents once or twice before.
2. 0:06AI agents are autonomous systems. They perceive their environment. They
3. 0:13make decisions based on those decisions. They can take actions. And this is all in
4. 0:20service of achieving a goal. But how do different agents talk to each
5. 0:26other to solve complex problems that a given agent can't solve by itself? Like, well, like
6. 0:33travel planning, you might need to integrate a bunch of agents together, like a
7. 0:39travel agent, and maybe that needs to integrate with a
8. 0:45flight agent and a hotel agent and an excursion agent
9. 0:52and so forth. Well, you could just build all of these agents yourself and then integrate them
10. 0:58with some custom code. But what if you want to use somebody else's hotel agent without knowing how
11. 1:04that agent communicates and how that agent works? It's not an easy task unless. Well, unless there
12. 1:10was a standard way for AI agents to work with each other. Something that allows
13. 1:16collaboration between agents, something that can handle authentication requirements, and something
14. 1:23that defines a common communication method as well. That is what the
15. 1:30A to A protocol, also known as the agent to
16. 1:36agent protocol, was designed to do. It was initially introduced by Google in April
17. 1:432025, and a to A is now open source housed by the Linux Foundation. So
18. 1:50how does it work? Well, let's start by talking about the core actors in A to A interactions. So
19. 1:57we have got here a user. Now that could be a human operator, or it could be an automated service.
20. 2:03And it's the user that initiates a request, or it sets a goal which is going to need the help of
21. 2:10one or more AI agents. So that request is received by
22. 2:16the client agent, and that acts on behalf of the
23. 2:23user and initiates requests to other agents. Those other agents, those
24. 2:30are called remote agents. I'm just going to draw one remote
25. 2:37agent here, but there could be a bunch of these in a mesh. But essentially these are
26. 2:43all AI agents. And in some scenarios, a given AI agent might be considered the client agent
27. 2:50making the cause. And in other cases, it might be considered the remote agent that's receiving them.
28. 2:56Now, the connection between the client agent and the remote
29. 3:02agent uses the A to A protocol. Oh, and you might also
30. 3:09see the client agent referred to as the A to A client, and the
31. 3:16remote agent referred to as the A to A server, if you see that.
32. 3:23Well, it just means the same thing. So let's get into how this all works by looking at
33. 3:29three stages. And we're going to start with number one which is discovery. So
34. 3:36how does the client agent find the remote agent and figure out what that agent
35. 3:43can actually do. Well it turns out that this remote agent actually publishes
36. 3:49something called an agent card. And that contains
37. 3:56basic information about the agent. So its identity, its capabilities, its skills. And it also
38. 4:03includes a service endpoint URL that enables the a two way communication and some authentication
39. 4:09requirements. And all of this takes the form of a JSON metadata document, which
40. 4:16is served on the agent's domain. Okay, so let's focus in on the agent interactions here. So the
41. 4:23client agent now knows how to find the remote agent. So now it is time for
42. 4:29stage two. That is authentication. Now that happens through security
43. 4:36scheme indicated in the agent card. So this is all based on
44. 4:43authentication based on security scheme. Now when the client agent has been successfully or
45. 4:50authenticated then the remote agent is responsible for authorization and
46. 4:57granting access control permissions. Now, with that all out of the way, we can move on to
47. 5:03stage three, which is communication. So in this case, the client agent is now ready to
48. 5:10send a task to the remote agent. An agent to agent communication, A
49. 5:17to A that uses the JSON RPC
50. 5:232.0 as the format for data exchange. And that is sent over
51. 5:30https. Now, the remote agent's job now is to
52. 5:36actually do the work. So it starts to process the task. And if
53. 5:43it requires more information, it can notify the client agent to say, hey, I need some more
54. 5:50information. So we might request more information here, which the client agent
55. 5:56then provides. It then sends back to us. And then once this remote agent has
56. 6:03actually completed its task, it can send a message to the
57. 6:09client along with any generated artifacts and an artifact
58. 6:16here is a tangible output that's generated by an agent during a task. So it could be a document or
59. 6:22an image or structured data. Now this communication that I've described so far is
60. 6:28request response. But some tasks might take the remote agent a little while to complete,
61. 6:35like when there's human interaction involved, or perhaps where there are external events. So for
62. 6:41long running tasks like those, if the agent card says that the remote agent supports streaming,
63. 6:47then we can use streaming with SSE. That's server send events, that
64. 6:54sends status updates about a given task from the remote agent to the client over an
65. 7:00open HTTP connection. So the agent agent protocol provides some pretty useful benefits when it
66. 7:07comes to discovery and authentication and a standardized communication, but also in the plus
67. 7:14column, I think we really need to also include privacy. The protocol treats
68. 7:21agentic AI as opaque agents, meaning the autonomous agents can collaborate without
69. 7:27actually having to reveal their inner workings, such as their internal memory or proprietary
70. 7:31logic, or any particular tool implementations that they use. It's pretty useful for preserving data
71. 7:37privacy and IP, and because A to A builds on established standards, these
72. 7:44are things that people are already using. And by that I'm talking about standards like
73. 7:50HTTP and JSON, RPC and then SSD
74. 7:58as well. Because of that, it makes it easier for enterprises to adopt the protocol. But look, A
75. 8:05to A is in its early stages. This is all pretty new stuff and things will
76. 8:11still improve. There are challenges still to overcome. There need to be improvements in
77. 8:16security and governance and performance tuning, just to name a few. But at its core, A to A
78. 8:22provides a way for AI agents to communicate over a trusted universal channel. It's a common
79. 8:28language for agent ecosystems, and that's a good thing to have, because I have a feeling that this
80. 8:34won't be the last time we talk about AI agents on this channel.