1
00:00:00,266 --> 00:00:01,600
All right.

2
00:00:01,600 --> 00:00:04,800
So today you're working with the alum.

3
00:00:05,500 --> 00:00:08,500
I have a new concept.

4
00:00:08,533 --> 00:00:08,800
Yeah.

5
00:00:08,800 --> 00:00:11,800
You were talking about feature owners,
right?

6
00:00:11,933 --> 00:00:12,300
Right.

7
00:00:12,300 --> 00:00:14,333
Which are people right now?

8
00:00:14,333 --> 00:00:16,000
No, but they're actors.

9
00:00:16,000 --> 00:00:16,766
They're actors.

10
00:00:16,766 --> 00:00:17,466
Okay.

11
00:00:17,466 --> 00:00:20,466
So actors could be an agent or a person.

12
00:00:20,933 --> 00:00:22,600
Okay. Right.

13
00:00:22,600 --> 00:00:25,366
And let's talk about

14
00:00:25,366 --> 00:00:29,366
the concept of a feature owner first
because I'm familiar with product owners.

15
00:00:29,400 --> 00:00:29,833
Yes, yes.

16
00:00:29,833 --> 00:00:32,833
Let me
let me give you the motivating example.

17
00:00:33,333 --> 00:00:36,333
I have a thing that with every feature

18
00:00:36,633 --> 00:00:40,800
request there's the invisible requirement
and everything else still works.

19
00:00:40,833 --> 00:00:43,033
Yeah. And this is the hardest requirement.

20
00:00:43,033 --> 00:00:46,900
And it's the requirement
that gets bigger and bigger and bigger,

21
00:00:46,933 --> 00:00:48,400
even though it's still invisible.

22
00:00:48,400 --> 00:00:52,300
As your app gets bigger
as everything else is a wider kind of.

23
00:00:52,300 --> 00:00:54,400
It's like the dark matter of of software.

24
00:00:54,400 --> 00:00:57,666
Yeah. The the feature interactions. Okay.

25
00:00:58,033 --> 00:00:59,733
Concrete example.

26
00:00:59,733 --> 00:01:03,433
I've got my little mtg deck shuffler.

27
00:01:03,433 --> 00:01:05,766
There's cards.

28
00:01:06,733 --> 00:01:10,000
Almost all magic cards have

29
00:01:10,300 --> 00:01:13,300
the same back
and then a face with stuff on them.

30
00:01:14,066 --> 00:01:18,133
But there's a few that have two meaningful
faces.

31
00:01:18,600 --> 00:01:21,700
They have like transform
this card under certain circumstances.

32
00:01:22,200 --> 00:01:24,633
So those are complicated.

33
00:01:24,633 --> 00:01:27,833
And early in the app development
I was like

34
00:01:27,833 --> 00:01:30,833
oh crap I need to handle two faced cards.

35
00:01:31,166 --> 00:01:34,166
And added flip card features.

36
00:01:34,600 --> 00:01:36,233
Okay.

37
00:01:36,233 --> 00:01:39,766
Many times
since then I've got to add a feature.

38
00:01:40,700 --> 00:01:43,900
And missed supporting two face cards.

39
00:01:43,933 --> 00:01:44,900
Okay.

40
00:01:44,900 --> 00:01:48,933
Concrete example I needed to add card

41
00:01:48,966 --> 00:01:52,000
type color identity mana cost.

42
00:01:52,000 --> 00:01:55,733
So some some properties of each card
I wanted to record those

43
00:01:56,833 --> 00:01:59,833
not just what's his name was its image.

44
00:02:00,066 --> 00:02:03,000
So I wanted to record those and I missed,

45
00:02:03,000 --> 00:02:04,400
two face cards.

46
00:02:04,400 --> 00:02:06,400
Yeah.

47
00:02:06,400 --> 00:02:10,300
So in real life,
what I've kind of always wanted on

48
00:02:10,300 --> 00:02:14,666
my team is every feature
we implement has a feature owner.

49
00:02:16,300 --> 00:02:17,966
In the past, it would have been a person.

50
00:02:17,966 --> 00:02:22,366
So an actor, 
who understands this feature.

51
00:02:22,766 --> 00:02:24,600
Who looks at every new incoming

52
00:02:24,600 --> 00:02:27,600
feature and says
does this interact with my feature.

53
00:02:28,466 --> 00:02:31,466
Make sure that supported, you know

54
00:02:32,100 --> 00:02:36,300
so that person has the knowledge
of where this feature came from.

55
00:02:36,300 --> 00:02:38,333
What it's for everywhere.

56
00:02:38,333 --> 00:02:41,233
It's currently implemented in the app.

57
00:02:41,233 --> 00:02:45,433
And the other things that it interacts
with I have that knowledge base.

58
00:02:46,000 --> 00:02:49,500
Their job is defined by what they know,
not by what they do.

59
00:02:49,500 --> 00:02:52,500
I think that's a lot of knowledge work.

60
00:02:52,666 --> 00:02:56,800
Okay, so the feature in this job,
in this case, I created one.

61
00:02:56,866 --> 00:03:02,266
Your job is to handle
two faced cards and the flip feature.

62
00:03:02,266 --> 00:03:05,066
Okay. So and when you say you created one.

63
00:03:05,066 --> 00:03:06,233
Good question.

64
00:03:06,233 --> 00:03:10,233
Of course, I created by, 
asking the AI to create it.

65
00:03:10,233 --> 00:03:11,000
So I'm asking

66
00:03:12,100 --> 00:03:15,100
and an AI agent, which is an actor.

67
00:03:15,700 --> 00:03:18,033
To create me another actor.

68
00:03:18,033 --> 00:03:19,700
Okay.

69
00:03:19,700 --> 00:03:22,533
With the knowledge base that I need.

70
00:03:22,533 --> 00:03:23,600
Okay.

71
00:03:23,600 --> 00:03:28,466
And whose job is to review
feature plans for interactions.

72
00:03:29,066 --> 00:03:32,066
And then later, come back and test

73
00:03:32,166 --> 00:03:35,666
and make sure that the
the flip feature still work.

74
00:03:35,833 --> 00:03:36,200
Okay.

75
00:03:36,200 --> 00:03:39,233
Now, so far, what you described so far,
that much sounds like

76
00:03:39,233 --> 00:03:42,833
just a secondary prompt, like a prompt
that you can summon up when you need to.

77
00:03:43,933 --> 00:03:46,133
So what makes it a feature owner?

78
00:03:46,133 --> 00:03:49,700
What makes it a feature owner is there's
a prompt, so that's its job description.

79
00:03:49,700 --> 00:03:51,933
You're a feature owner.
This is what you do. Okay.

80
00:03:51,933 --> 00:03:53,933
And it's knowledge base.

81
00:03:53,933 --> 00:03:55,000
What it. Okay.

82
00:03:55,000 --> 00:03:58,000
So it has a directory in the repo.

83
00:03:58,300 --> 00:03:59,133
It owns a directory.

84
00:03:59,133 --> 00:04:00,233
It owns a directory, okay.

85
00:04:00,233 --> 00:04:02,200
Where it can store its knowledge.

86
00:04:02,200 --> 00:04:04,533
It's writing the features of markdown
files.

87
00:04:04,533 --> 00:04:05,866
Yeah. Okay. Yeah.

88
00:04:07,366 --> 00:04:10,366
So these
two things kind of make the actor,

89
00:04:10,766 --> 00:04:14,833
and then the actor can't help
until it's invoked, right?

90
00:04:15,233 --> 00:04:16,100
That's the thing.

91
00:04:16,100 --> 00:04:19,100
how does it become that person on the team
Who's like, wait a minute.

92
00:04:20,033 --> 00:04:21,100
What about. Yeah.

93
00:04:21,100 --> 00:04:24,000
How's it going to impact,
two phase cards?

94
00:04:24,000 --> 00:04:25,033
Yeah. And you're planning meeting?

95
00:04:25,033 --> 00:04:28,133
Yes. And, 
what about after you flip the card?

96
00:04:28,133 --> 00:04:30,366
Does it still work?

97
00:04:30,366 --> 00:04:32,433
Yes, exactly.

98
00:04:32,433 --> 00:04:35,100
So what we need to do is get that person
in the planning meeting.

99
00:04:35,100 --> 00:04:37,900
The actor. Sorry. In the planning meeting.

100
00:04:37,900 --> 00:04:40,033
And that winds up.

101
00:04:40,033 --> 00:04:43,233
I mean I just asked Claude how to do this
and Claude did it and it worked.

102
00:04:44,000 --> 00:04:47,266
That winds up in the agent
R&D Claude, DMD,

103
00:04:47,500 --> 00:04:50,733
with its process
for implementing features.

104
00:04:51,300 --> 00:04:53,100
There's a plan step. Okay.

105
00:04:53,100 --> 00:04:56,666
And the plan step includes
invoking all the feature owners.

106
00:04:57,600 --> 00:05:00,400
So the flip card features
the library grouping features

107
00:05:00,400 --> 00:05:01,966
that all the ones that I've created.

108
00:05:01,966 --> 00:05:07,033
So when you're interacting,
with an agent, as you're developing,

109
00:05:07,366 --> 00:05:11,833
what I'm gathering is you're primarily
interacting with kind of a project manager

110
00:05:11,833 --> 00:05:15,466
or a, a team manager process.

111
00:05:15,500 --> 00:05:20,100
Ya, that has a, as part of
its prompt has a process of like,

112
00:05:20,200 --> 00:05:24,566
delegate to this, sub prompt or sub agent.

113
00:05:24,566 --> 00:05:25,666
And then this is something we.

114
00:05:25,666 --> 00:05:27,600
Have to do more
and more of these days, right?

115
00:05:27,600 --> 00:05:30,366
So that the main agent
that we're interacting with.

116
00:05:30,366 --> 00:05:33,033
It doesn't get its context
to full of details.

117
00:05:33,033 --> 00:05:34,033
Right? Right.

118
00:05:34,033 --> 00:05:36,766
At an extreme,
there's stuff like client teams,

119
00:05:37,966 --> 00:05:40,966
which is incredibly high overhead.

120
00:05:41,000 --> 00:05:44,266
But it it actively spins up
a bunch of agent actors.

121
00:05:45,566 --> 00:05:47,700
And keeps them

122
00:05:47,700 --> 00:05:49,200
open.

123
00:05:49,200 --> 00:05:52,200
The feature owner that I have is simpler.

124
00:05:52,266 --> 00:05:54,200
It's a call and response.

125
00:05:54,200 --> 00:05:56,333
It's a hey, here's the plan.

126
00:05:56,333 --> 00:05:59,333
Do you see any interactions
we should worry about to

127
00:06:00,200 --> 00:06:03,700
invoke the actor with its system prompt,

128
00:06:04,033 --> 00:06:07,033
and then the system prompt includes,
here's your knowledge base.

129
00:06:07,400 --> 00:06:10,000
And then the invocation of that actor

130
00:06:10,000 --> 00:06:12,800
has two outputs.

131
00:06:12,800 --> 00:06:15,300
Just like any function just like us. Okay.

132
00:06:15,300 --> 00:06:18,200
There's whatever turns

133
00:06:18,200 --> 00:06:21,133
and there's the side effect
of who it becomes.

134
00:06:21,133 --> 00:06:23,566
Right. So who it becomes is it's updated.

135
00:06:23,566 --> 00:06:24,533
It's knowledge base

136
00:06:24,533 --> 00:06:28,333
that the next time that actor is invoked
it has different knowledge.

137
00:06:28,366 --> 00:06:29,066
Okay.

138
00:06:29,066 --> 00:06:32,600
And then there's the output of actually
watch out for this.

139
00:06:33,500 --> 00:06:35,166
Make sure you do this. Okay.

140
00:06:36,166 --> 00:06:37,666
Make sure you store

141
00:06:37,666 --> 00:06:42,766
the card type and the man of value
and stuff for the back face of the card.

142
00:06:43,200 --> 00:06:44,800
Okay.

143
00:06:44,800 --> 00:06:45,500
That's interesting.

144
00:06:45,500 --> 00:06:49,100
And then that is somehow carried forward
into whatever like design,

145
00:06:49,100 --> 00:06:53,000
detailed design sub-agent you have.

146
00:06:53,133 --> 00:06:55,000
Yeah. Whatever agent it might be.

147
00:06:55,000 --> 00:06:58,000
The main agent is doing the planning,

148
00:06:58,066 --> 00:07:00,800
gets this information
back and incorporates it into the plan.

149
00:07:00,800 --> 00:07:02,533
Okay. And proceed.

150
00:07:02,533 --> 00:07:05,533
So in a sense, I mean,
from my perspective, it's it's like,

151
00:07:05,566 --> 00:07:08,566
just like it would invoke a,

152
00:07:08,666 --> 00:07:13,266
a suite of tests or just like
it would invoke, linting tool.

153
00:07:13,433 --> 00:07:18,533
It also invokes some agents and sees what,
what output they have.

154
00:07:18,533 --> 00:07:21,233
And you can lend to the planning phase.

155
00:07:21,233 --> 00:07:23,333
So yeah, it's kind of plan lint in a way.

156
00:07:23,333 --> 00:07:27,433
I mean, it's not a, predefined
set of rules that don't change in the way

157
00:07:27,433 --> 00:07:30,766
most linting tools
are. But it's still like

158
00:07:32,266 --> 00:07:33,133
it's still invoking in

159
00:07:33,133 --> 00:07:36,466
a kind of, in the same sense as a, as a,
as it would a linting tool.

160
00:07:36,900 --> 00:07:39,600
Yes, yes. It's a don't forget.

161
00:07:39,600 --> 00:07:44,833
And, and the effect of this is that the,
everything else that still works

162
00:07:44,833 --> 00:07:47,833
in the code base is divided up

163
00:07:48,000 --> 00:07:51,033
into little modules of knowledge.

164
00:07:51,600 --> 00:07:57,066
And there's steps, a growing list
of steps as our feature base grows.

165
00:07:58,300 --> 00:08:00,100
Of does it affect this.

166
00:08:00,100 --> 00:08:01,433
Does it affect this.

167
00:08:01,433 --> 00:08:05,633
And the effect of that
is that instead of having to understand

168
00:08:05,633 --> 00:08:09,900
everything all at once,
we understand it in various pieces.

169
00:08:10,466 --> 00:08:13,600
And, and,
and all that knowledge comes together

170
00:08:13,800 --> 00:08:18,033
in the main agent who's the project lead
for this new feature.

171
00:08:18,066 --> 00:08:19,233
Whatever.

172
00:08:19,233 --> 00:08:23,733
And then at the end of that new feature
implementation you create a feature owner

173
00:08:24,666 --> 00:08:27,733
like okay
project lead encapsulate what you know.

174
00:08:28,966 --> 00:08:30,400
Into this new

175
00:08:30,400 --> 00:08:33,500
knowledge base and write yourself
a feature on our instructions.

176
00:08:34,233 --> 00:08:36,900
So that next time you are in charge

177
00:08:36,900 --> 00:08:39,900
of making sure your feature still works.

178
00:08:40,833 --> 00:08:41,133
Yeah.

179
00:08:41,133 --> 00:08:44,133
And I guess eventually you kind of have to

180
00:08:44,466 --> 00:08:47,466
maybe factor out the job of

181
00:08:47,833 --> 00:08:50,166
notifying all the feature owners.

182
00:08:50,166 --> 00:08:55,266
Maybe my, my biggest, you know, the thing
that I observe over and over again

183
00:08:55,266 --> 00:08:58,866
is that any the further back
something is in the list of things to do.

184
00:08:59,900 --> 00:09:02,966
The more it falls off the brain of
of LLMs.

185
00:09:03,233 --> 00:09:05,466
I was hearing it.

186
00:09:05,466 --> 00:09:07,200
It's part of the list of things to do.

187
00:09:07,200 --> 00:09:10,200
And it might be the same agent
that does the implementation.

188
00:09:10,566 --> 00:09:13,800
And my experience is that,
like if you have something

189
00:09:13,800 --> 00:09:15,600
doing implementation,
it's going to forget about what.

190
00:09:15,600 --> 00:09:16,133
Oh, wow.

191
00:09:16,133 --> 00:09:17,866
The agent helps doing the planning. Yeah.

192
00:09:17,866 --> 00:09:20,066
The agent that's doing the
planning for the new feature.

193
00:09:21,500 --> 00:09:22,300
Because they really are they

194
00:09:22,300 --> 00:09:25,300
get very focused and they,
they they lose things.

195
00:09:25,366 --> 00:09:25,700
Right.

196
00:09:25,700 --> 00:09:30,633
But but part of the the team's concept is
don't do the implementation in the agent.

197
00:09:30,733 --> 00:09:32,566
That makes. Sense.

198
00:09:32,566 --> 00:09:35,566
So I'm not much of a developer actor.

199
00:09:36,000 --> 00:09:39,466
So much of what we do,
I feel like when we are working

200
00:09:39,466 --> 00:09:43,200
more with agents is
we are managing their attention.

201
00:09:44,266 --> 00:09:45,333
Absolutely.

202
00:09:45,333 --> 00:09:47,466
Just like we do with ourselves. Yeah.

203
00:09:47,466 --> 00:09:52,000
But but the fun part is, I, I've often
wished in the past that I could

204
00:09:52,000 --> 00:09:57,100
fork myself and do multiple things
and with with agents we absolutely can.

205
00:09:57,600 --> 00:10:00,033
So this part of myself
that knows about the project

206
00:10:00,033 --> 00:10:03,600
that the feature that was just implemented
fork it off into a feature owner.

207
00:10:04,300 --> 00:10:05,733
Come back to that one later.

208
00:10:07,666 --> 00:10:08,166
Yeah.

209
00:10:08,166 --> 00:10:08,700
Yeah.

210
00:10:08,700 --> 00:10:11,700
We are we're

211
00:10:13,333 --> 00:10:16,333
we have many units of attention now.

212
00:10:16,833 --> 00:10:20,000
And so we can chunk or delegate.

213
00:10:20,766 --> 00:10:23,766
We have to delegate the knowledge.

214
00:10:24,333 --> 00:10:25,133
Right. Yeah.

215
00:10:25,133 --> 00:10:28,133
And we don't have the, the UN teams.

216
00:10:28,166 --> 00:10:31,166
We can't just keep people around forever
just to know

217
00:10:32,100 --> 00:10:34,300
a thing really well.

218
00:10:34,300 --> 00:10:37,200
And in agents we absolutely can.

219
00:10:37,200 --> 00:10:37,833
Yeah.

220
00:10:37,833 --> 00:10:40,833
I'm interested
to see if you run into like a new,

221
00:10:42,866 --> 00:10:45,566
I don't know, accumulative snowball effect
where.

222
00:10:45,566 --> 00:10:47,933
I get to the new failure modes
as fast as we can.

223
00:10:47,933 --> 00:10:50,933
That's the game. Yeah.

224
00:10:51,566 --> 00:10:53,900
I'm already imagining that
if I were writing the prompts

225
00:10:53,900 --> 00:10:56,266
for these things,
I would have to be very adamant about.

226
00:10:56,266 --> 00:11:01,566
Like, unless you see a clear interaction,
just say no problem.

227
00:11:01,600 --> 00:11:06,633
And because I can totally see them
decide each one, deciding

228
00:11:06,666 --> 00:11:10,100
to pontificate
about its own particular area of expertise

229
00:11:10,100 --> 00:11:14,466
and all the possible implications
that then you think, no, no, no, no.

230
00:11:14,666 --> 00:11:15,433
Right.

231
00:11:15,433 --> 00:11:19,266
We have to have agents that want to go
back. To.

232
00:11:20,166 --> 00:11:20,866
sorting the magic cards.

233
00:11:23,066 --> 00:11:23,466
Yeah.

234
00:11:23,466 --> 00:11:23,800
Yeah.