LangGraph: Intro — Transcript

Full transcript

1. 0:00Hello everyone.
2. 0:02Today I want to talk about LangGraph, a
3. 0:05new library that we're releasing.
4. 0:07So LangGraph builds on top of LangChain
5. 0:10and makes it really easy to create
6. 0:12agents and agent runtimes.
7. 0:15So what exactly is an agent and an agent
8. 0:17runtime?
9. 0:18So in LangChain, we define an agent as a
10. 0:20system powered by a language model that
11. 0:22decides what actions to take.
12. 0:25There's then an agent runtime that
13. 0:26basically runs that agent in a loop,
14. 0:29calls that agent, decides what actions
15. 0:31to take, then takes that action, and
16. 0:32then get records the observation, and
17. 0:35then passes that back in and starts the
18. 0:36loop over again, and continues going
19. 0:38through that loop until the agent
20. 0:40decides that it is finished.
21. 0:43So we've made it easy over the past few
22. 0:45months to customize agents in LangChain
23. 0:47with LangChain Expression Language, and
24. 0:49I'd recommend checking those out if you
25. 0:50haven't.
26. 0:52What we're doing with LangGraph is we're
27. 0:53making it easy to customize the agent
28. 0:55runtime. So previously, the agent
29. 0:58runtime was always the agent executor
30. 1:00class. And this was basically something
31. 1:02that ran in a loop and and called tools
32. 1:05in a specific way and and handled errors
33. 1:07in a specific way, and that was great,
34. 1:10but it was just one way to create this
35. 1:11runtime. And so we want to create more
36. 1:14ways to create this runtime and ways to
37. 1:16create them more flexibly and
38. 1:18dynamically.
39. 1:19And a key part of this runtime is the
40. 1:21ability to add cycles. So you know, the
41. 1:24whole point of the agent runtime is
42. 1:26running this agent, this LLM-powered
43. 1:29agent, in a loop, and so you need to be
44. 1:31able to have cycles. And LangChain
45. 1:32Expression Language and other you know,
46. 1:35DAG-like frameworks are non, they're not
47. 1:39cyclical.
48. 1:40And so that's why we're introducing
49. 1:41LangGraph for.
50. 1:42It's it's a way to create these cyclical
51. 1:45agent runtimes.
52. 1:47So that's hopefully an overview of what
53. 1:50LangGraph is and and why we created it.
54. 1:53The rest of the videos in this series
55. 1:54are going I'm going on two of the main
56. 1:57agent runtimes that we've added to
57. 1:59start. So, first we've added agent
58. 2:01executor very similar to the agent
59. 2:04executor in LangChain. So, we've
60. 2:06recreated that agent executor with
61. 2:08LangGraph. And then we've added a chat
62. 2:10agent executor. The chat agent executor
63. 2:13takes in a list of messages and then
64. 2:15represents the agent state just as a
65. 2:17list of messages and so it returns a
66. 2:19list of messages as well.
67. 2:21The reason we did this is that a lot of
68. 2:22the newer models are chat-based models
69. 2:25and they intrinsically represent
70. 2:26function calling
71. 2:28as as basically parameters as part of a
72. 2:30message and then they also have function
73. 2:32responses being separate type of
74. 2:34message. And so representing this agent
75. 2:37state as a list of messages is very
76. 2:39natural for these types of models. And
77. 2:41so we've added a separate agent executor
78. 2:43particularly for that.
79. 2:45Then I'm going to highlight a bunch of
80. 2:46different ways that you can modify the
81. 2:48base agent executors to do things like
82. 2:50add human in the loops,
83. 2:52force specific tool to be called first,
84. 2:55and other exciting things like that.
85. 2:57Hope you enjoy.