1
00:00:00,06 --> 00:00:03,03
- [Narrator] In the last video, we set up our environment,

2
00:00:03,03 --> 00:00:07,06
but now let's talk about what an AI agent actually is.

3
00:00:07,06 --> 00:00:09,07
And you can think of an agent as a program

4
00:00:09,07 --> 00:00:13,01
that understands your goals, makes a plan to achieve them,

5
00:00:13,01 --> 00:00:15,01
and then takes action.

6
00:00:15,01 --> 00:00:17,01
At the heart of the agent is this core loop.

7
00:00:17,01 --> 00:00:19,07
It consists of these four key steps,

8
00:00:19,07 --> 00:00:23,08
understanding your intent, making a plan, using tools,

9
00:00:23,08 --> 00:00:25,07
and then reflecting on what it did.

10
00:00:25,07 --> 00:00:27,04
So let's talk through these steps.

11
00:00:27,04 --> 00:00:29,07
First of all is understanding intent,

12
00:00:29,07 --> 00:00:32,03
and this all starts with you, the user.

13
00:00:32,03 --> 00:00:35,00
You're going to give the agent a task like finding a place

14
00:00:35,00 --> 00:00:36,08
for hike near where I am.

15
00:00:36,08 --> 00:00:39,07
The first thing that the agent has to do is understand

16
00:00:39,07 --> 00:00:42,00
what it is that you're asking for.

17
00:00:42,00 --> 00:00:45,01
It needs to figure out what your intent is.

18
00:00:45,01 --> 00:00:47,06
And in this case, the intent is to go for a hike

19
00:00:47,06 --> 00:00:49,04
at a nearby national park.

20
00:00:49,04 --> 00:00:51,06
Now you can use a large language model to do that,

21
00:00:51,06 --> 00:00:54,00
so there's many different ways that I could say this,

22
00:00:54,00 --> 00:00:55,06
I want to hike, I want to go for a hike,

23
00:00:55,06 --> 00:00:57,02
please help me with a hike.

24
00:00:57,02 --> 00:00:59,04
And the nice thing there about the large language model

25
00:00:59,04 --> 00:01:02,05
is that instead of doing the traditional natural language

26
00:01:02,05 --> 00:01:04,08
processing to try and figure out what you said,

27
00:01:04,08 --> 00:01:07,03
the LLM can actually artificially understand

28
00:01:07,03 --> 00:01:10,08
that on your behalf and turn that into an intent.

29
00:01:10,08 --> 00:01:13,06
So now once the agent has derived the intent

30
00:01:13,06 --> 00:01:15,03
from what you said you want to do,

31
00:01:15,03 --> 00:01:17,02
now it needs to make a plan.

32
00:01:17,02 --> 00:01:19,07
Now this plan means it breaks down the task

33
00:01:19,07 --> 00:01:22,03
into smaller manageable steps.

34
00:01:22,03 --> 00:01:23,08
So for our hiking agent,

35
00:01:23,08 --> 00:01:25,09
that plan might look something like this.

36
00:01:25,09 --> 00:01:28,01
Well, first we need to know where the user is,

37
00:01:28,01 --> 00:01:29,05
what's their location.

38
00:01:29,05 --> 00:01:32,04
Secondly, we need to get the weather at that location.

39
00:01:32,04 --> 00:01:34,01
Maybe today is not a good day for a hike

40
00:01:34,01 --> 00:01:36,06
because it's raining or snowing, that kind of thing.

41
00:01:36,06 --> 00:01:38,05
So if it is suitable, let's continue.

42
00:01:38,05 --> 00:01:41,09
And to continue would be we'll find nearby national parks,

43
00:01:41,09 --> 00:01:45,02
and we'll get more details about them and then we'll parse

44
00:01:45,02 --> 00:01:48,09
the details about those parks to get hiking suggestions,

45
00:01:48,09 --> 00:01:53,05
and then give the user some opinionated suggestions

46
00:01:53,05 --> 00:01:56,00
based on the weather, based on the parks

47
00:01:56,00 --> 00:01:57,06
and those types of things.

48
00:01:57,06 --> 00:02:00,01
Now all of that involves using tools.

49
00:02:00,01 --> 00:02:03,03
So in order for the agent to be able to make a plan,

50
00:02:03,03 --> 00:02:05,09
it needs to know what tools are available to it.

51
00:02:05,09 --> 00:02:08,03
And once it knows the tools are available to it,

52
00:02:08,03 --> 00:02:10,07
then it can as part of the planning process,

53
00:02:10,07 --> 00:02:12,03
use those tools.

54
00:02:12,03 --> 00:02:14,08
Now typically those tools are going to be APIs,

55
00:02:14,08 --> 00:02:15,06
of course, right?

56
00:02:15,06 --> 00:02:18,07
So the agent needs to know how to call an API for weather,

57
00:02:18,07 --> 00:02:20,09
how to call the API for location,

58
00:02:20,09 --> 00:02:22,07
and then even like the national parks,

59
00:02:22,07 --> 00:02:25,07
how to call the API for that and to parse the results.

60
00:02:25,07 --> 00:02:28,05
And this is where agents get really, really powerful.

61
00:02:28,05 --> 00:02:31,07
Those tools are what the agent uses to intersect,

62
00:02:31,07 --> 00:02:34,04
and interact with the outside world.

63
00:02:34,04 --> 00:02:36,00
Now, a tool, like I said, could be anything

64
00:02:36,00 --> 00:02:39,02
from a simple function like convert Celsius to Fahrenheit,

65
00:02:39,02 --> 00:02:41,02
or it could be a full blown API like searching

66
00:02:41,02 --> 00:02:42,06
for national parks.

67
00:02:42,06 --> 00:02:45,01
And our hiking agent is going to use those tools

68
00:02:45,01 --> 00:02:47,06
that I've mentioned, understand where the user is,

69
00:02:47,06 --> 00:02:49,02
fetch the weather where they are,

70
00:02:49,02 --> 00:02:51,08
and then search for national parks if appropriate,

71
00:02:51,08 --> 00:02:54,07
to be able to hike based on the current weather.

72
00:02:54,07 --> 00:02:56,05
Once it's done all of this kind of thing,

73
00:02:56,05 --> 00:02:58,05
it then needs to reflect,

74
00:02:58,05 --> 00:03:01,04
it will take look at the results of its actions and ask,

75
00:03:01,04 --> 00:03:03,06
well, did I accomplish this goal?

76
00:03:03,06 --> 00:03:06,08
LLMs are well known for hallucination, for example,

77
00:03:06,08 --> 00:03:09,03
and this is one of those steps where you can take a look

78
00:03:09,03 --> 00:03:12,01
at what has happened, look at the user's intent,

79
00:03:12,01 --> 00:03:15,03
look at the details from the tools, look at the suggestions,

80
00:03:15,03 --> 00:03:18,09
and say, hey, I really did think I accomplished the goal.

81
00:03:18,09 --> 00:03:22,00
Or if not, then go back and try again,

82
00:03:22,00 --> 00:03:24,05
or maybe ask the user for more details.

83
00:03:24,05 --> 00:03:26,02
Are you okay to hike in the snow?

84
00:03:26,02 --> 00:03:27,05
Those type of things.

85
00:03:27,05 --> 00:03:29,08
So that's where the reflection part of it

86
00:03:29,08 --> 00:03:31,02
will often give you the results

87
00:03:31,02 --> 00:03:33,09
because the results are good, but if they're not good,

88
00:03:33,09 --> 00:03:35,07
a really good agent will go back,

89
00:03:35,07 --> 00:03:38,06
take a look at its results and go back and maybe replan

90
00:03:38,06 --> 00:03:41,04
and maybe indeed as part of replanning also go back

91
00:03:41,04 --> 00:03:44,06
to maybe get extra intent from the user.

92
00:03:44,06 --> 00:03:47,06
So to recap, this is how you should be thinking

93
00:03:47,06 --> 00:03:49,07
when you're building agents to break it down

94
00:03:49,07 --> 00:03:51,02
into these four steps.

95
00:03:51,02 --> 00:03:54,03
To understand the user's intent, to plan,

96
00:03:54,03 --> 00:03:57,00
to act on that plan, and then to reflect.

97
00:03:57,00 --> 00:03:59,08
And this is the basic loop that all agents will follow.

98
00:03:59,08 --> 00:04:02,03
So next up, we're finally going to get our hands dirty

99
00:04:02,03 --> 00:04:05,00
and we'll write our very first agent.