WEBVTT

1
00:00:00.600 --> 00:00:04.745
This is a practical blueprint for building a personal AI agent.

2
00:00:04.745 --> 00:00:08.193
It covers the agent loop, twelve building steps,

3
00:00:08.193 --> 00:00:11.583
the full stack and the order to build it in.

4
00:00:11.583 --> 00:00:15.507
Then it maps every piece to tools the agency already runs,

5
00:00:15.507 --> 00:00:17.272
and shows where that can grow.

6
00:00:18.818 --> 00:00:23.218
A personal AI agent is an AI system that understands goals,

7
00:00:23.218 --> 00:00:27.421
uses tools, finds the right information,

8
00:00:27.421 --> 00:00:30.091
takes action and remembers context.

9
00:00:30.974 --> 00:00:34.631
It works through tasks with as much or as little independence as

10
00:00:34.631 --> 00:00:35.536
you allow.

11
00:00:37.063 --> 00:00:39.873
Every agent runs the same basic loop.

12
00:00:39.873 --> 00:00:43.216
It starts with a goal and works to understand it.

13
00:00:43.216 --> 00:00:47.547
It makes a plan, chooses a tool and takes action.

14
00:00:47.547 --> 00:00:50.240
Then it checks the result, adjusts,

15
00:00:50.240 --> 00:00:52.249
and repeats until the job is done.

16
00:00:53.792 --> 00:00:57.252
Step one: define its job.

17
00:00:57.252 --> 00:01:00.456
Start with one specific problem, not ten.

18
00:01:01.664 --> 00:01:06.458
Good first jobs include research, summarizing documents,

19
00:01:06.458 --> 00:01:10.754
preparing meeting briefs, handling repetitive tasks,

20
00:01:10.754 --> 00:01:16.176
analyzing data, organizing information and writing first drafts.

21
00:01:17.705 --> 00:01:20.863
Step two: choose the AI model.

22
00:01:21.745 --> 00:01:24.172
The model is the agent's brain.

23
00:01:24.172 --> 00:01:29.594
Compare models on reasoning, accuracy, speed and cost,

24
00:01:29.594 --> 00:01:32.090
on how much context they can hold,

25
00:01:32.090 --> 00:01:34.238
and on how well they use tools.

26
00:01:35.767 --> 00:01:41.061
Step three: write clear instructions in five parts.

27
00:01:41.061 --> 00:01:44.440
The role says who the agent is.

28
00:01:44.440 --> 00:01:47.110
The goal says what it must accomplish.

29
00:01:47.992 --> 00:01:51.011
The workflow says how it works.

30
00:01:51.011 --> 00:01:54.343
The rules say what it must always follow,

31
00:01:54.343 --> 00:01:57.454
and the limits say what it must never do.

32
00:01:59.001 --> 00:02:01.799
Step four: give it knowledge.

33
00:02:02.832 --> 00:02:07.500
Connect what it needs to read: documents, PDFs, notes,

34
00:02:07.500 --> 00:02:11.749
standard operating procedures, project files, databases,

35
00:02:11.749 --> 00:02:13.862
knowledge bases and cloud storage.

36
00:02:15.391 --> 00:02:21.126
Step five: give it tools, because tools let an agent act.

37
00:02:21.126 --> 00:02:26.955
Common tools are web search, file search, code execution, email,

38
00:02:26.955 --> 00:02:33.317
calendar, databases, APIs, browser tools and custom functions.

39
00:02:34.863 --> 00:02:39.995
Step six: connect your apps through APIs, function calling,

40
00:02:39.995 --> 00:02:45.103
webhooks, MCP servers and automation platforms.

41
00:02:45.103 --> 00:02:48.807
This is what turns a chatbot into a system that works with your

42
00:02:48.807 --> 00:02:50.002
other software.

43
00:02:51.540 --> 00:02:54.001
Step seven: add memory.

44
00:02:55.035 --> 00:02:58.239
Working memory holds the current task.

45
00:02:58.239 --> 00:03:01.165
Conversation memory holds past exchanges.

46
00:03:02.198 --> 00:03:04.590
External memory lives in documents,

47
00:03:04.590 --> 00:03:06.505
databases and knowledge stores.

48
00:03:07.388 --> 00:03:11.138
Long-term memory keeps the important facts that must last.

49
00:03:12.684 --> 00:03:15.052
Step eight: add planning.

50
00:03:15.935 --> 00:03:20.439
For complex work, the agent breaks a goal into a plan,

51
00:03:20.439 --> 00:03:26.105
completes one step, checks the result, chooses the next step,

52
00:03:26.105 --> 00:03:27.313
and repeats.

53
00:03:28.839 --> 00:03:32.902
Step nine: add guardrails to protect your systems

54
00:03:32.902 --> 00:03:33.796
and information.

55
00:03:34.830 --> 00:03:38.638
Use permission controls, output validation,

56
00:03:38.638 --> 00:03:43.723
approval requirements, action limits, input validation,

57
00:03:43.723 --> 00:03:46.579
tool restrictions and sandboxing.

58
00:03:48.128 --> 00:03:52.412
Step ten: keep humans in control.

59
00:03:52.412 --> 00:03:54.792
Automate low-risk actions.

60
00:03:54.792 --> 00:03:57.811
Automate and monitor medium-risk actions.

61
00:03:57.811 --> 00:04:01.886
Require approval for high-risk actions, like sending email,

62
00:04:01.886 --> 00:04:06.565
deleting files, making purchases or changing important settings.

63
00:04:08.097 --> 00:04:10.523
Step eleven: automate it,

64
00:04:10.523 --> 00:04:13.577
but only once the workflow runs reliably.

65
00:04:13.577 --> 00:04:17.919
Add triggers such as every morning, every week, a new email,

66
00:04:17.919 --> 00:04:22.261
a new file, a form submission, a calendar event,

67
00:04:22.261 --> 00:04:25.187
a database update or an API event.

68
00:04:26.733 --> 00:04:30.506
Step twelve: test and improve.

69
00:04:30.506 --> 00:04:36.764
Measure tool selection, reliability, task success, accuracy,

70
00:04:36.764 --> 00:04:40.886
speed, cost, safety and failure rate.

71
00:04:41.768 --> 00:04:45.588
Use logs and tracing to see what the agent actually did.

72
00:04:47.120 --> 00:04:52.054
Put together, the personal AI agent stack has eight layers: the

73
00:04:52.054 --> 00:04:57.894
AI model, instructions, knowledge, tools, memory, automation,

74
00:04:57.894 --> 00:04:59.566
guardrails and monitoring.

75
00:05:01.106 --> 00:05:02.697
Build it in this order.

76
00:05:02.697 --> 00:05:04.519
Pick one task.

77
00:05:04.519 --> 00:05:06.307
Choose the model.

78
00:05:06.307 --> 00:05:08.478
Write the instructions.

79
00:05:08.478 --> 00:05:10.034
Add knowledge.

80
00:05:10.034 --> 00:05:11.880
Connect tools.

81
00:05:11.880 --> 00:05:13.622
Add memory.

82
00:05:13.622 --> 00:05:15.386
Add guardrails.

83
00:05:15.386 --> 00:05:16.826
Test.

84
00:05:16.826 --> 00:05:17.592
Automate.

85
00:05:18.475 --> 00:05:19.636
Then improve.

86
00:05:21.180 --> 00:05:25.220
This direction opens new service lines: a one-job agent built

87
00:05:25.220 --> 00:05:30.166
for a client, a voice front desk, daily operations briefs,

88
00:05:30.166 --> 00:05:34.392
a knowledge base build, a guardrails and approval review,

89
00:05:34.392 --> 00:05:38.572
an automation and monitoring plan, hands-on workshops,

90
00:05:38.572 --> 00:05:40.917
and narrated explainers like this one.

91
00:05:42.454 --> 00:05:43.987
Start small.

92
00:05:43.987 --> 00:05:48.120
Pick one job, build it in order, keep a person in control,

93
00:05:48.120 --> 00:05:49.223
and measure everything.

94
00:05:50.256 --> 00:05:52.160
The full documentation,

95
00:05:52.160 --> 00:05:55.922
the mapping table and a build checklist are on this page.