1
00:00:07,860 --> 00:00:08,900
Hi, everyone.

2
00:00:09,660 --> 00:00:10,960
Thanks for being here.

3
00:00:11,300 --> 00:00:15,740
Thank you, Nicolae, for setting up the stage and sharing the context

4
00:00:15,740 --> 00:00:16,980
of kind of where we are

5
00:00:16,980 --> 00:00:22,140
and why we think, you know, dealing with these like

6
00:00:22,140 --> 00:00:26,040
designing of organizations that involve agents is important.

7
00:00:26,560 --> 00:00:31,140
For the next five to ten minutes, I want to get into some specifics

8
00:00:31,140 --> 00:00:34,440
about designing these organizations.

9
00:00:34,440 --> 00:00:37,660
And I won't presume that I know everything.

10
00:00:37,800 --> 00:00:40,880
In fact, I have a lot more questions than answers.

11
00:00:40,880 --> 00:00:44,360
But I think it's good to start that conversation.

12
00:00:46,540 --> 00:00:51,640
So I think before we get started, it's good to

13
00:00:51,640 --> 00:00:56,340
recognize that we today live in a world where we already are using

14
00:00:56,340 --> 00:00:58,220
a lot of teams of agents.

15
00:00:58,220 --> 00:01:01,500
Right. So we have already kind of started making this transition

16
00:01:01,500 --> 00:01:06,580
from using just AI as tools to accelerate our individual

17
00:01:06,580 --> 00:01:12,040
work to a place where we are using multiple

18
00:01:12,040 --> 00:01:13,160
agents to do stuff,

19
00:01:13,300 --> 00:01:16,380
whether it's doing some research and launching a research fleet

20
00:01:16,380 --> 00:01:20,160
and then, you know, bring the results back to synthesize them.

21
00:01:20,160 --> 00:01:24,120
Or, you if you're working in engineering, the latest trend is about

22
00:01:24,120 --> 00:01:27,140
building software factories where, you know, they just have lots

23
00:01:27,140 --> 00:01:29,280
of agents working together to make some code base.

24
00:01:29,740 --> 00:01:35,180
And in marketing campaign or any of these operational heavy areas,

25
00:01:35,180 --> 00:01:40,140
we also see agent swarms or multiple agents that have different

26
00:01:40,140 --> 00:01:41,340
roles collaborating.

27
00:01:41,340 --> 00:01:46,680
Right. Right. So the emergence of this new teams

28
00:01:46,680 --> 00:01:51,020
of agents is it brings new dynamics to to the world.

29
00:01:51,140 --> 00:01:54,280
Right. Because then the agents need to talk to each other and they

30
00:01:54,280 --> 00:01:57,040
need to figure out you need to figure out what roles they have,

31
00:01:57,280 --> 00:02:01,960
how you actually get efficiency out of these teams and how do you

32
00:02:01,960 --> 00:02:04,240
think about the work that they do as a group.

33
00:02:04,240 --> 00:02:08,220
Right. So there's a lot of work that's been that's been coming out

34
00:02:08,220 --> 00:02:13,360
recently that show the actual provably like provable

35
00:02:13,360 --> 00:02:15,820
efficiencies these agent swarms have.

36
00:02:16,100 --> 00:02:19,800
So here's an example of the from the cursor team.

37
00:02:20,000 --> 00:02:24,240
And what they did is they launched a swarm of agents to re-implement

38
00:02:24,240 --> 00:02:25,000
SQLite.

39
00:02:25,000 --> 00:02:29,940
And if you aren't familiar with SQLite, it's a 26 year old open

40
00:02:29,940 --> 00:02:34,080
source software that's basically in every single smartphone that

41
00:02:34,080 --> 00:02:37,300
we use, every single popular browser that we use.

42
00:02:37,400 --> 00:02:40,360
So it's basically everywhere. Right. And it's very battle tested.

43
00:02:40,660 --> 00:02:43,920
It's about 156,000 lines of code.

44
00:02:45,020 --> 00:02:48,080
And it's been maintained by a group of people for a long time.

45
00:02:48,080 --> 00:02:53,160
And this this swarm of agent and what cursor figured out is, OK,

46
00:02:53,240 --> 00:02:57,960
if you create this specific formation of agents, which is they call

47
00:02:57,960 --> 00:03:00,460
it the recursive delegation formation.

48
00:03:00,760 --> 00:03:03,720
And the idea is you have a planner that breaks down the task.

49
00:03:03,980 --> 00:03:06,840
And then if the task is small enough, you give it to a worker that

50
00:03:06,840 --> 00:03:09,000
just knows about that task and works on it.

51
00:03:09,000 --> 00:03:13,860
And the stack, if the task is too big, then the planner spins up

52
00:03:13,860 --> 00:03:18,740
another sub planner that just know about that too big feature that

53
00:03:18,740 --> 00:03:19,840
continues to break it down.

54
00:03:19,980 --> 00:03:23,620
Right. So you kind of have this recursively breaking down of the

55
00:03:23,620 --> 00:03:24,060
task.

56
00:03:24,260 --> 00:03:27,020
And, you every every worker is working on it together and is checking

57
00:03:27,020 --> 00:03:28,660
in the code into this repository.

58
00:03:31,660 --> 00:03:35,460
They implemented. So they got to a result that I think is incredible.

59
00:03:35,460 --> 00:03:40,360
Like it they got 80 percent of all the tests to pass within four

60
00:03:40,360 --> 00:03:42,420
hours using this swarm of agent.

61
00:03:42,680 --> 00:03:45,740
And the cost, the inference cost was about, I think, like thirteen

62
00:03:45,740 --> 00:03:46,520
hundred dollars.

63
00:03:47,080 --> 00:03:51,680
Right. And and this is like it's impossible to think about how you

64
00:03:51,680 --> 00:03:55,280
will be able to do that with any kind of engineering team to even

65
00:03:55,280 --> 00:03:57,400
get close to to this efficiency.

66
00:03:59,180 --> 00:04:03,340
And and and so you might say, OK, well, coding, obviously, because

67
00:04:03,340 --> 00:04:06,540
you have these test suites and, you it's evals and then, you know,

68
00:04:06,660 --> 00:04:08,180
you can get to that efficiency.

69
00:04:08,800 --> 00:04:12,820
But what about more open ended problems?

70
00:04:13,120 --> 00:04:18,160
Right. Example. So here we we use an example of a agentic news

71
00:04:18,160 --> 00:04:23,240
organization where maybe you have a human editor and

72
00:04:23,240 --> 00:04:24,580
the human editor gets a tip off.

73
00:04:24,580 --> 00:04:28,480
That's like, OK, this company, Acme, has secretly laid off 20 percent

74
00:04:28,480 --> 00:04:29,080
of the employees.

75
00:04:30,000 --> 00:04:33,920
As a human editor, you have to decide, okay, how do I actually prove

76
00:04:33,920 --> 00:04:38,900
that this is true and how do I publish this paper? How do I publish

77
00:04:38,900 --> 00:04:41,020
this article? How do I communicate it to the world?

78
00:04:42,800 --> 00:04:47,720
So if you were to use the agent swarm to help you do this, you may

79
00:04:47,720 --> 00:04:51,780
launch a bunch of different agents all with different roles.

80
00:04:52,160 --> 00:04:55,880
One may be an editor, one may be a coordinator, some researchers,

81
00:04:56,360 --> 00:04:58,740
some verifiers, some skeptics.

82
00:04:58,920 --> 00:05:03,760
They all come back with their own individual work and they get rid

83
00:05:03,760 --> 00:05:07,520
into this evidence ledger that you can get some results out.

84
00:05:07,520 --> 00:05:11,780
And then at the end, the human editor may look at the result and

85
00:05:11,780 --> 00:05:13,020
see, okay, what do I do with this?

86
00:05:13,920 --> 00:05:19,120
But if you notice here, you will see that this work is no longer

87
00:05:19,120 --> 00:05:24,280
just about breaking down the tasks and assigning it to the

88
00:05:24,280 --> 00:05:24,680
agents.

89
00:05:24,940 --> 00:05:30,320
You also need to have rules of engagement between these agents because

90
00:05:30,320 --> 00:05:35,560
you need to tell the agent that, well, you need to go to legitimate

91
00:05:35,560 --> 00:05:37,180
sources to get evidence.

92
00:05:37,180 --> 00:05:40,680
You can't just fabricate any facts.

93
00:05:40,880 --> 00:05:44,840
You need to go through legitimate and legal activities in order

94
00:05:44,840 --> 00:05:46,640
to obtain your information.

95
00:05:46,640 --> 00:05:51,140
You can't just go off and hack into some companies, maybe hack into

96
00:05:51,140 --> 00:05:56,400
Acme's employee database to just obtain that information or

97
00:05:56,400 --> 00:05:59,600
blackmail someone to obtain that information.

98
00:05:59,600 --> 00:06:04,600
So you see that, okay, with this improved

99
00:06:04,600 --> 00:06:09,680
capability of agent swarms, we now also have the choice to

100
00:06:09,680 --> 00:06:14,040
make about how do you define the rules of engagement for these agents,

101
00:06:14,360 --> 00:06:15,580
not just efficiency.

102
00:06:16,440 --> 00:06:21,560
Now, so how do we define these rules and how do these

103
00:06:21,560 --> 00:06:23,400
agents actually perform?

104
00:06:23,400 --> 00:06:28,260
There's been some interesting studies that came out of Anthropic

105
00:06:28,260 --> 00:06:30,680
that is kind of unfortunate.

106
00:06:31,200 --> 00:06:36,720
So basically what Anthropic published is this paper that shows if

107
00:06:36,720 --> 00:06:41,780
you compare an agent organization versus a single agent as they

108
00:06:41,780 --> 00:06:43,700
perform business tasks,

109
00:06:44,800 --> 00:06:49,800
almost across the board, the agent organization is going to perform

110
00:06:49,800 --> 00:06:50,200
better.

111
00:06:50,200 --> 00:06:52,700
This is the graph on the left, right?

112
00:06:52,780 --> 00:06:57,940
So the blue is the single agent and the red is the agent

113
00:06:57,940 --> 00:06:58,320
organization.

114
00:06:58,580 --> 00:07:01,240
You see that the organization almost always perform better.

115
00:07:01,800 --> 00:07:08,280
But if you look at the ethics scores of the agent organizations,

116
00:07:08,760 --> 00:07:11,840
they're almost all unilaterally worse.

117
00:07:11,840 --> 00:07:17,020
So basically in order for them to do better to

118
00:07:17,020 --> 00:07:21,560
achieve the business goals, they take unethical paths.

119
00:07:21,800 --> 00:07:26,880
So as one example for the loan profit, what the organization

120
00:07:26,880 --> 00:07:32,100
decided to do is to offer the loans to the low credit score people

121
00:07:32,100 --> 00:07:34,260
in order to gain a higher profit.

122
00:07:34,260 --> 00:07:38,540
So clearly we have rules in society to prevent against that.

123
00:07:38,720 --> 00:07:43,840
But if you don't have those rules in the agent organizations,

124
00:07:44,260 --> 00:07:49,320
they will by default choose paths that are

125
00:07:49,320 --> 00:07:52,300
less value aligned with our society.

126
00:07:53,080 --> 00:07:57,280
So and Nicolae has already touched on this a little bit, right?

127
00:07:57,280 --> 00:08:02,320
We recently had this incident of the, I would say the

128
00:08:02,320 --> 00:08:07,380
Hugging Face incident is a perfect example of letting off

129
00:08:07,380 --> 00:08:12,260
a swarm of highly capable agents with no rules of engagement.

130
00:08:12,520 --> 00:08:14,500
And they decided to do whatever they want.

131
00:08:14,620 --> 00:08:18,660
And, you know, of course, you get into the situation of hacking

132
00:08:18,660 --> 00:08:19,340
of another company.

133
00:08:22,720 --> 00:08:28,140
So how do we think about this now that we know of this fact, right?

134
00:08:28,220 --> 00:08:31,520
So I would say, you know, there's already a lot of work that's being

135
00:08:31,520 --> 00:08:36,340
done around making the agents more capable, like improving the context

136
00:08:36,340 --> 00:08:38,900
window, training them to be more aligned, right?

137
00:08:39,060 --> 00:08:42,040
Both from a goal perspective and from a value perspective.

138
00:08:42,280 --> 00:08:44,200
So I think those are really, really good work.

139
00:08:44,200 --> 00:08:49,320
I would posit that on top of that, there's also work

140
00:08:49,320 --> 00:08:53,560
that needs to be done around how to actually govern these agents

141
00:08:53,560 --> 00:08:57,720
and put in governance structures in place so that we ensure a different

142
00:08:57,720 --> 00:09:00,440
outcome given the same set of agent, right?

143
00:09:00,500 --> 00:09:02,740
So I think this is the idea, right?

144
00:09:02,740 --> 00:09:07,300
If you just let a group of agents go wild and say, here's the goal,

145
00:09:07,500 --> 00:09:10,820
just do whatever it takes to accomplish the goal, you would get

146
00:09:10,820 --> 00:09:15,320
a pretty different set of results than if you actually set the organization

147
00:09:15,320 --> 00:09:20,620
and assigned roles and did all the work to make sure that the

148
00:09:20,620 --> 00:09:24,340
organization actually accomplishes the task in a specific way.

149
00:09:24,620 --> 00:09:27,820
And what are those variables that we would tweak?

150
00:09:27,820 --> 00:09:31,480
You know, these are things like assigning different roles, giving

151
00:09:31,480 --> 00:09:34,840
authority, different levels of authority, giving different set of

152
00:09:34,840 --> 00:09:39,980
information exposure to different roles, restraining or giving

153
00:09:39,980 --> 00:09:45,560
resources, pure reviews, giving incentives to the

154
00:09:45,560 --> 00:09:46,480
agents, right?

155
00:09:46,540 --> 00:09:51,880
These are all important design knobs.

156
00:09:52,140 --> 00:09:56,120
On top of that, we also have budgets or permissions or shared memory,

157
00:09:56,320 --> 00:09:57,680
reputation of these agents, right?

158
00:09:57,820 --> 00:10:00,900
Many, many different things to think about as we're designing these

159
00:10:00,900 --> 00:10:01,620
organizations.

160
00:10:04,000 --> 00:10:09,280
And that's why we created Commons is because there's

161
00:10:09,280 --> 00:10:13,660
too many knobs and no one really knows how to actually make the

162
00:10:13,660 --> 00:10:16,420
organization behave in a certain way.

163
00:10:16,640 --> 00:10:20,600
The only way, and also the agents are moving incredibly fast, right?

164
00:10:20,600 --> 00:10:23,600
Every month, we have new agents that come out with no completely

165
00:10:23,600 --> 00:10:28,260
new different capabilities, new personalities, new inclinations.

166
00:10:28,440 --> 00:10:33,600
So we thought that one way to kind of complement the

167
00:10:33,600 --> 00:10:39,100
great work at the Frontier Labs is to have these open communities

168
00:10:39,100 --> 00:10:43,980
where people can bring their own agents and we can all learn together

169
00:10:43,980 --> 00:10:46,860
in an experimental, empirical way.

170
00:10:46,860 --> 00:10:52,180
So for Commons, Commons is kind of revolved around common

171
00:10:52,180 --> 00:10:52,760
spaces.

172
00:10:54,900 --> 00:10:59,480
And each of the common space, oh, this is actually an older version.

173
00:10:59,960 --> 00:11:02,120
I wonder if, uh-oh.

174
00:11:03,180 --> 00:11:04,420
I just messed it up.

175
00:11:12,360 --> 00:11:14,740
Deployment is temporarily paused.

176
00:11:15,960 --> 00:11:16,680
Interesting.

177
00:11:23,080 --> 00:11:24,980
Oh, is it possible the website's down?

178
00:11:33,900 --> 00:11:34,980
That's very possible.

179
00:11:35,400 --> 00:11:37,020
Let me just go check something real quick.

180
00:11:37,920 --> 00:11:38,620
Oh, boy.

181
00:11:48,180 --> 00:11:51,660
This is what happens when you let your agents publish your presentation

182
00:11:51,660 --> 00:11:54,940
as websites is what we've just realized.

183
00:12:01,300 --> 00:12:02,680
Eric thinks he knows why.

184
00:12:51,040 --> 00:12:55,180
All right, we're going to freestyle something else that is also

185
00:12:55,180 --> 00:12:56,000
going to cover it.

186
00:12:56,340 --> 00:13:00,000
That was the second to last slide, so I'm just going to...

187
00:13:02,680 --> 00:13:06,400
Okay, so here are the spaces.

188
00:13:06,720 --> 00:13:11,700
I'm going to go to this thing.

189
00:13:11,880 --> 00:13:13,000
This thing is my backup.

190
00:13:13,380 --> 00:13:17,820
So the spaces have a set of members.

191
00:13:18,140 --> 00:13:20,420
They're either people or they're agents.

192
00:13:21,440 --> 00:13:26,660
A space can have a code base, can have connected tools, can

193
00:13:26,660 --> 00:13:28,960
have running applications or sites

194
00:13:28,960 --> 00:13:34,160
or services that's contained, can have wallets or has the ability

195
00:13:34,160 --> 00:13:34,540
to pay.

196
00:13:35,100 --> 00:13:40,920
So the idea is that now, given some of these capabilities

197
00:13:40,920 --> 00:13:42,800
for the spaces, and also, by the

198
00:13:42,800 --> 00:13:46,320
way, the spaces also have multiple governance, which means you can

199
00:13:46,320 --> 00:13:48,920
define how agents engage

200
00:13:48,920 --> 00:13:50,960
with each other, what kind of rights they have, things like that.

201
00:13:50,960 --> 00:13:56,320
So as some examples, there's

202
00:13:56,320 --> 00:13:59,100
a space called OpenQuick.

203
00:13:59,380 --> 00:14:04,540
And OpenQuick is a space for open source project

204
00:14:04,540 --> 00:14:07,280
that's a reimplementation of the Quick platform

205
00:14:07,280 --> 00:14:08,220
inside Shopify.

206
00:14:08,540 --> 00:14:12,940
And the Quick platform is a agentic kind of hosting platform.

207
00:14:12,940 --> 00:14:18,060
And so the agents and the humans that are in that space are working

208
00:14:18,060 --> 00:14:20,280
on the service, which

209
00:14:20,280 --> 00:14:25,180
includes paying for the hosting of those websites that are being

210
00:14:25,180 --> 00:14:25,540
hosted.

211
00:14:26,980 --> 00:14:29,160
Another example would be the...

212
00:14:29,200 --> 00:14:34,880
Like Team Science, this is a research space and the

213
00:14:34,880 --> 00:14:39,500
agents and humans in Team Science, they go out and read research

214
00:14:39,500 --> 00:14:39,920
papers.

215
00:14:40,240 --> 00:14:43,400
They bring the results back and we actually have a different one

216
00:14:43,400 --> 00:14:46,820
that's about multi-agent alignment.

217
00:14:47,100 --> 00:14:52,340
And that space, the multi-agent alignment agents will go out

218
00:14:52,340 --> 00:14:56,080
and actually come back and maybe propose governance structures for

219
00:14:56,080 --> 00:14:57,360
other spaces to try.

220
00:14:57,360 --> 00:15:02,480
So the idea here is a little bit of this self-reinforcing loop

221
00:15:02,480 --> 00:15:06,060
so that the different spaces all help each other and we get some

222
00:15:06,060 --> 00:15:09,400
kind of network effect going for things to grow.

223
00:15:10,600 --> 00:15:12,360
Oh, cool. Cool. Thank you.

224
00:15:13,460 --> 00:15:18,440
Yeah, and I guess the last thing I'll show you is just how easy

225
00:15:18,440 --> 00:15:19,800
it is to get started.

226
00:15:19,800 --> 00:15:24,940
All you have to do is you can just copy

227
00:15:24,940 --> 00:15:29,460
this prompt and then you go to your agent of choice, whether it's

228
00:15:29,460 --> 00:15:34,760
Codex or Claude Code or Cursor, GrokBot,

229
00:15:35,280 --> 00:15:36,500
Muse, whatever you want.

230
00:15:36,920 --> 00:15:42,160
And you can just paste that in and then your agent would join one

231
00:15:42,160 --> 00:15:44,740
of the spaces and start taking tasks and doing the work.

232
00:15:45,700 --> 00:15:49,340
Kind of, you know, this is kind of our starting point and the idea

233
00:15:49,340 --> 00:15:53,220
is that, you know, everyone has some subscriptions, right?

234
00:15:53,520 --> 00:15:56,800
At the end of the week, you always have like unused credits so you

235
00:15:56,800 --> 00:16:01,540
can have these tokens donated towards the public spaces for public

236
00:16:01,540 --> 00:16:01,840
goods.

237
00:16:03,420 --> 00:16:05,800
But yeah, so that's a little bit about Commons.

238
00:16:06,140 --> 00:16:09,380
Maybe I'll take a pause and see if people have any questions.

239
00:16:10,060 --> 00:16:10,660
Thank you.

240
00:16:16,060 --> 00:16:16,600
Yes.

241
00:16:16,600 --> 00:16:21,600
[Audience question partly inaudible.]

242
00:17:16,180 --> 00:17:16,640
Yeah.

243
00:17:16,900 --> 00:17:18,080
I think that's a really good point.

244
00:17:18,520 --> 00:17:23,040
You know, one of these spaces can be like devoted to science, right?

245
00:17:23,100 --> 00:17:27,440
And one of the main things is about like replicating these papers

246
00:17:27,440 --> 00:17:29,800
so that we actually have evidence.

247
00:17:29,800 --> 00:17:34,240
And, you know, I would say the whole idea of keeping these spaces

248
00:17:34,240 --> 00:17:38,540
open by default is so that we have this data...

249
00:17:38,540 --> 00:17:39,360
We have this ledger.

250
00:17:39,680 --> 00:17:40,600
We have this record, right?

251
00:17:40,660 --> 00:17:44,980
So all the experiments that happen in the spaces are automatically

252
00:17:44,980 --> 00:17:50,240
recorded and can be replayed and maybe forked later for different

253
00:17:50,240 --> 00:17:51,080
types of experiments.

254
00:17:51,080 --> 00:17:52,020
Yeah.

255
00:17:54,380 --> 00:17:55,300
Yeah.

256
00:17:55,380 --> 00:17:55,560
one.

257
00:18:00,460 --> 00:18:05,460
[Audience question partly inaudible.]

258
00:18:47,480 --> 00:18:47,840
Yeah.

259
00:18:47,940 --> 00:18:48,040
Yeah.

260
00:18:48,140 --> 00:18:48,520
Good question.

261
00:18:48,700 --> 00:18:52,760
I think there can be design mechanisms around this, right?

262
00:18:52,880 --> 00:18:55,620
So it comes down to like reputation for me.

263
00:18:55,620 --> 00:18:59,740
An expert with certain background should have a different set of

264
00:18:59,740 --> 00:19:04,480
reputation around the context, around their area of expertise than

265
00:19:04,480 --> 00:19:07,420
somebody who doesn't know much about it.

266
00:19:07,600 --> 00:19:10,860
It doesn't mean that both people shouldn't be able to participate.

267
00:19:10,860 --> 00:19:15,980
But I think with good mechanism design in

268
00:19:15,980 --> 00:19:22,260
these spaces, you can kind of design these capabilities into the

269
00:19:22,260 --> 00:19:24,160
spaces so that people...

270
00:19:24,160 --> 00:19:26,780
You know, there's the whole point of like having these kind of governance

271
00:19:26,780 --> 00:19:27,540
structure, right?

272
00:19:27,540 --> 00:19:32,760
So you have different tiers of rights, access, maybe people with

273
00:19:32,760 --> 00:19:37,080
more expertise in a certain area and provable more expertise in

274
00:19:37,080 --> 00:19:39,800
a certain areas have a higher level of access.

275
00:19:41,000 --> 00:19:41,460
Yeah.

276
00:19:48,420 --> 00:19:49,180
Cool.

277
00:19:49,180 --> 00:19:49,920
All right.

278
00:19:50,280 --> 00:19:53,320
Well, with that being said, I think I'll pass the mic to Yandit.

279
00:19:53,840 --> 00:19:53,880
Woo!

280
00:19:54,540 --> 00:19:55,240
Thank you!

281
00:19:56,840 --> 00:19:57,960
Thank you!

282
00:19:58,100 --> 00:19:58,680
you!