1
00:00:00,021 --> 00:00:20,901
Good morning, everybody. Thank you for joining us today. It is such an honor to be sharing this stage with so many great speakers. So yeah, let's get right into it. My name's Aman. I lead AI at Arca.

2
00:00:20,901 --> 00:00:26,821
and today I'm going to talk to you about ARCA's Long Horizon agents that run our business.

3
00:00:28,061 --> 00:00:31,521
ARCA is an AI native services firm.

4
00:00:31,841 --> 00:00:39,781
We provide our clients with financial services like financial planning, tax planning, asset management, investment advice,

5
00:00:40,421 --> 00:00:42,361
and we serve a range of clients.

6
00:00:42,581 --> 00:00:49,381
Some of them are like technology executives that have amassed a lot of wealth and need a real plan to retire early.

7
00:00:49,381 --> 00:00:54,221
but also that family in New Hampshire that made their first million dollars

8
00:00:54,221 --> 00:00:56,781
and want to buy that house finally.

9
00:00:59,121 --> 00:01:00,801
I said we were AI native.

10
00:01:01,361 --> 00:01:06,041
There's lots of definitions of this term these days, very overused.

11
00:01:06,221 --> 00:01:12,001
What we mean by AI native is that agents run our business.

12
00:01:12,401 --> 00:01:16,541
Not that they assist the business from some co-pilot corner.

13
00:01:16,541 --> 00:01:22,661
They execute the decisions that shape the outcomes that our business has for our clients.

14
00:01:24,701 --> 00:01:28,441
Work in this industry, it's interesting to know what it looks like.

15
00:01:28,581 --> 00:01:33,741
In financial services, the work is more long horizon and less long running.

16
00:01:33,961 --> 00:01:41,401
And what I mean by that is this is less agents grinding tokens for like hours or days on end, right?

17
00:01:41,581 --> 00:01:45,821
This is more one client goal that spans 10 days.

18
00:01:45,821 --> 00:01:55,741
It's very stop-start. There's many actors in between, many decision points, many handoffs, and ultimately that goal lands like 10 days later.

19
00:01:55,901 --> 00:01:58,461
That is the kind of work that we are working on.

20
00:01:59,441 --> 00:02:04,821
And so here's an example. Jane's our client, fairly wealthy, but has a big tax bill.

21
00:02:05,221 --> 00:02:11,481
Oh. Has a big tax bill in April 15th, right?

22
00:02:11,481 --> 00:02:21,021
So we're going to help Jane figure out what is the most tax-optimal way for this cash to be raised to pay the tax bill.

23
00:02:21,601 --> 00:02:28,261
So there's the advisor that will do all of the statement analysis and the analysis of how to make this tax optimal.

24
00:02:28,641 --> 00:02:33,141
There's the trader that would actually liquidate the assets.

25
00:02:34,221 --> 00:02:39,921
assets themselves will take days to settle, and then the ops people will come in and trigger the ACH.

26
00:02:39,921 --> 00:02:42,461
that will itself take days to land.

27
00:02:42,701 --> 00:02:45,561
So this is kind of the work I'm talking about in this industry,

28
00:02:46,201 --> 00:02:48,641
which takes days to settle.

29
00:02:49,061 --> 00:02:52,561
And of course, if you do this at scale for several, several clients,

30
00:02:52,861 --> 00:02:56,421
the amount of time our firm takes to do this sort of work exponentiates, right?

31
00:02:56,941 --> 00:03:00,541
So at ARCA, we took a different approach.

32
00:03:01,301 --> 00:03:04,601
It's helpful to know that ARCA is not a SaaS company.

33
00:03:05,701 --> 00:03:09,581
We are not selling software to other businesses, to other firms.

34
00:03:09,581 --> 00:03:11,201
We are the firm.

35
00:03:11,541 --> 00:03:15,781
We're an applied AI company that happens to be running a services business in this industry.

36
00:03:16,481 --> 00:03:23,201
And so the approach we took was, all right, we're going to have agents own this whole long horizon outcome end to end.

37
00:03:24,121 --> 00:03:28,441
And for that, we gave them a couple of key components I'll give you a sneak peek of.

38
00:03:28,721 --> 00:03:34,221
We gave them signals from the external world that keeps them really aware of what is happening.

39
00:03:34,221 --> 00:03:36,541
So that email will be read by agents.

40
00:03:36,541 --> 00:03:40,441
We gave them all of our operational context on what to do about that email.

41
00:03:40,641 --> 00:03:42,981
What does this industry do with the situation of taxes?

42
00:03:43,861 --> 00:03:49,381
We gave them a harness that gives them all of the tools and the guardrails,

43
00:03:49,601 --> 00:03:54,641
and more importantly, persistence of memory over the long horizon period

44
00:03:54,641 --> 00:03:55,981
to carry that work out.

45
00:03:56,721 --> 00:04:00,281
And, of course, we gave them the ability to collaborate with people.

46
00:04:02,421 --> 00:04:07,361
So that work with agents compresses a lot,

47
00:04:07,361 --> 00:04:12,301
because all we did was we took humans out of the critical path of this work.

48
00:04:12,421 --> 00:04:16,301
Nobody has to babysit this goal over days.

49
00:04:16,541 --> 00:04:17,841
Agents are on it.

50
00:04:17,961 --> 00:04:20,581
In this talk, I'd love to describe to you how we actually do that.

51
00:04:23,021 --> 00:04:24,641
Let's follow the journey of this email.

52
00:04:26,641 --> 00:04:29,221
Jane owes $100,000 for taxes.

53
00:04:29,221 --> 00:04:35,621
this email, much like many other signals in our platform,

54
00:04:35,821 --> 00:04:38,041
will be read by background agents.

55
00:04:38,241 --> 00:04:40,201
So there's phone calls our firm is making,

56
00:04:40,461 --> 00:04:41,681
there's documents that are coming in,

57
00:04:41,781 --> 00:04:42,841
there's meetings being had,

58
00:04:43,521 --> 00:04:45,281
there's events coming in from our custodian,

59
00:04:45,561 --> 00:04:47,441
there's market movement that's happening.

60
00:04:47,601 --> 00:04:49,881
These are all signals that our background agents

61
00:04:49,881 --> 00:04:53,481
are continuously evaluating to progress work.

62
00:04:53,561 --> 00:04:55,201
So this email is one of those signals.

63
00:04:55,701 --> 00:04:58,001
And when the signal is ingested by agents,

64
00:04:58,141 --> 00:04:59,001
they actually get to the work.

65
00:04:59,221 --> 00:05:04,761
right? We have a harness that gives them the knowledge to know what to do about the situation.

66
00:05:05,121 --> 00:05:10,341
They have deep client memory, so they know, hey, let's never touch Jane's Apple position,

67
00:05:10,421 --> 00:05:15,801
because she said that in a call like eight months ago, right? We, of course, have a sandbox for all

68
00:05:15,801 --> 00:05:21,881
of the financial computation that needs to happen for this analysis. But largely, I think harnesses

69
00:05:21,881 --> 00:05:27,201
today are coalescing on a standard shape. So it's more interesting what happens outside of

70
00:05:27,201 --> 00:05:28,121
this harness.

71
00:05:30,781 --> 00:05:33,661
So, after they've done the work, they produce this plan.

72
00:05:34,021 --> 00:05:37,461
This plan by itself cannot be executed without people.

73
00:05:37,681 --> 00:05:42,541
So, in our product, we build a space, agents build a space, where they put out this plan,

74
00:05:42,621 --> 00:05:47,121
and you can chat with the agent to kind of know what's happening with the plan.

75
00:05:47,281 --> 00:05:50,261
There's multiple people that can be chatting with the agent.

76
00:05:50,581 --> 00:05:55,861
And more importantly, this space is where also they pull people in when they need people

77
00:05:55,861 --> 00:05:56,821
to execute some work.

78
00:05:57,201 --> 00:06:01,341
So here in this case, we would approve the plan and the email out to Jane.

79
00:06:02,821 --> 00:06:06,281
And then there's this whole dance of, has Jane replied?

80
00:06:06,901 --> 00:06:09,881
What does her answer mean for the plan that we had?

81
00:06:10,741 --> 00:06:13,821
And then all of these other questions that will have to be answered

82
00:06:13,821 --> 00:06:16,061
throughout this whole long-horizon work.

83
00:06:16,881 --> 00:06:21,321
At Arca, we were like, of course, we're not going to have people babysitting it.

84
00:06:21,341 --> 00:06:24,961
So we let our agents watch the world so people don't have to.

85
00:06:24,961 --> 00:06:30,681
We give them tools to be able to register interest in anything happening in our world.

86
00:06:30,801 --> 00:06:33,761
So they're able to say, let me know when Jane responds.

87
00:06:34,321 --> 00:06:36,321
Let me know when that ACH lands.

88
00:06:36,921 --> 00:06:40,901
And this is powered by our signal system in the background,

89
00:06:40,901 --> 00:06:46,441
where agents are routing signals to agents that are interested in it.

90
00:06:46,901 --> 00:06:49,481
And of course, if the signal never lands,

91
00:06:50,321 --> 00:06:53,961
agents can wake themselves up on a schedule as well

92
00:06:53,961 --> 00:06:55,201
to keep track of work.

93
00:06:57,421 --> 00:07:01,541
But then, of course, the real world is very messy, right?

94
00:07:01,801 --> 00:07:03,201
People are waiting for that email,

95
00:07:03,261 --> 00:07:06,341
but Jay never responds to that email.

96
00:07:07,101 --> 00:07:09,261
Say we meet her in person,

97
00:07:09,281 --> 00:07:12,161
and we learn some information that affects the plan.

98
00:07:12,561 --> 00:07:13,961
So the person goes in,

99
00:07:14,041 --> 00:07:15,981
and they add this information into our UI.

100
00:07:17,581 --> 00:07:21,001
But this act of adding information or doing work in our UI

101
00:07:21,001 --> 00:07:24,561
is also a signal that will wake up an agent.

102
00:07:24,841 --> 00:07:27,761
In fact, when people work in our product,

103
00:07:27,961 --> 00:07:29,921
when they reschedule a calendar event,

104
00:07:30,661 --> 00:07:31,541
that's a signal.

105
00:07:31,921 --> 00:07:34,721
When they mark a task complete, that is a signal.

106
00:07:35,201 --> 00:07:39,101
So this system lets us be both aware

107
00:07:39,101 --> 00:07:41,001
of everything happening in the external world

108
00:07:41,001 --> 00:07:43,381
and the internal world as work is progressing

109
00:07:43,381 --> 00:07:45,121
so our agents can participate in work.

110
00:07:48,461 --> 00:07:51,881
and of course agents are communicating with each other as well

111
00:07:51,881 --> 00:07:53,961
so this signal read by a background agent

112
00:07:53,961 --> 00:07:56,981
we let it notify another agent about this

113
00:07:56,981 --> 00:07:58,981
so that it can update the space to say

114
00:07:58,981 --> 00:08:02,561
hey, we just need $80,000 instead of the $100,000

115
00:08:02,561 --> 00:08:05,221
so the work has changed because of these signals

116
00:08:07,281 --> 00:08:09,181
and the last piece of the story is that

117
00:08:09,181 --> 00:08:11,721
agents are able to pull in the right people at the right time

118
00:08:13,101 --> 00:08:16,201
So the agent on the left has read operational context

119
00:08:16,201 --> 00:08:18,561
to be like, okay, the next step here

120
00:08:18,561 --> 00:08:22,121
is to pull in the person that trades these assets.

121
00:08:22,401 --> 00:08:24,701
And so that person now gets a task

122
00:08:24,701 --> 00:08:27,701
and gets pulled in to participate in this long horizon work.

123
00:08:29,721 --> 00:08:31,801
What we've done overall, if you take a step back,

124
00:08:31,821 --> 00:08:34,201
is we've given our agents the primitives

125
00:08:34,201 --> 00:08:37,101
for driving long horizon work.

126
00:08:37,301 --> 00:08:38,721
What do people do at work, right?

127
00:08:38,881 --> 00:08:40,941
They need to stay aware of what is happening,

128
00:08:40,941 --> 00:08:44,961
so we've built this very intense signal system for our agents to be aware.

129
00:08:45,381 --> 00:08:47,321
They need to know what to do next,

130
00:08:47,361 --> 00:08:49,261
so we've given them all this operational context.

131
00:08:49,841 --> 00:08:51,901
They need to do the work, so it's our harness

132
00:08:51,901 --> 00:08:54,661
that lets them actually perform the work and keep track of it.

133
00:08:54,881 --> 00:08:59,081
It's long persistence memory systems that hold the state

134
00:09:00,481 --> 00:09:02,421
and bring in the right people at the right time.

135
00:09:06,101 --> 00:09:08,921
This is a pretty simple idea on the face of it,

136
00:09:08,921 --> 00:09:11,261
but there's a lot of emergent complexity.

137
00:09:11,861 --> 00:09:13,541
You can imagine at any point,

138
00:09:13,721 --> 00:09:15,521
there's hundreds of these signals coming in,

139
00:09:15,781 --> 00:09:20,281
and tasks are being added on people's plates,

140
00:09:20,421 --> 00:09:21,201
they're being pulled in,

141
00:09:21,561 --> 00:09:23,481
background agents are communicating with each other.

142
00:09:24,661 --> 00:09:26,041
So over the last six months,

143
00:09:26,281 --> 00:09:28,961
we've been working on getting this system to be really good.

144
00:09:29,721 --> 00:09:31,861
We've been closing all of our context gaps

145
00:09:32,641 --> 00:09:34,141
so that agents are fully aware.

146
00:09:34,761 --> 00:09:37,581
We have been working on the precision of tasks,

147
00:09:37,581 --> 00:09:40,001
So pulling in the right person at the right time.

148
00:09:40,701 --> 00:09:44,761
Work that needs to be tracked for the future is tracked in the agent's memory.

149
00:09:45,181 --> 00:09:47,681
Work that is doable now becomes a task, right?

150
00:09:47,881 --> 00:09:52,161
So we've been working on our ability to hold state over long horizons as well.

151
00:09:52,641 --> 00:09:56,141
And because we are the firm and not just a SaaS company,

152
00:09:56,421 --> 00:09:59,881
we've been changing and standardizing how work happens in the

153
00:10:00,021 --> 00:10:01,201
firm in the first place.

154
00:10:01,521 --> 00:10:13,801
And operationalizing that context into our agent state. And because these are our colleagues that use our product, we collect rich reasoning from them that become evals for our system to improve.

155
00:10:16,281 --> 00:10:24,361
So I guess there's a question to be asked here, like, why do this? What does this unlock? What is the takeaway here?

156
00:10:25,241 --> 00:10:34,881
For us, this is our whole thesis, is that we want our platform to be so AI native that agents drive the business.

157
00:10:35,201 --> 00:10:37,061
Agents run the business.

158
00:10:37,321 --> 00:10:39,621
They are making the decisions that shape our outcomes.

159
00:10:40,421 --> 00:10:48,001
And so at ARCA, when people log in, they see a list of tasks that have been prepared for them by agents.

160
00:10:48,681 --> 00:10:50,241
They've already read that email.

161
00:10:50,241 --> 00:10:52,841
They've looked at the document that came in.

162
00:10:52,841 --> 00:10:54,781
they've done like 90% of the work,

163
00:10:54,981 --> 00:10:58,841
and they're pulling you in just when it's time for you to act.

164
00:10:59,881 --> 00:11:04,401
So building this system lets us get to a point where at Arca,

165
00:11:05,541 --> 00:11:09,241
agents prompt people more than people are prompting agents,

166
00:11:09,261 --> 00:11:13,721
which lets our agents drive the work that our firm does.

167
00:11:14,221 --> 00:11:17,121
And because agents are driving the work the firm does,

168
00:11:17,921 --> 00:11:19,381
agents can drive the business.

169
00:11:20,241 --> 00:11:21,281
Let me give you an example.

170
00:11:21,281 --> 00:11:23,261
Remember that email that Jane sent us?

171
00:11:23,781 --> 00:11:25,441
This is a bit of a failure, right?

172
00:11:25,541 --> 00:11:29,741
Because we should have just been proactive about this in the first place, right?

173
00:11:30,201 --> 00:11:31,921
So Jane should never have to send this email.

174
00:11:32,441 --> 00:11:37,701
So say at, you know, our AI native business, we take on this new goal.

175
00:11:37,941 --> 00:11:40,961
Our new goal is proactivity.

176
00:11:42,601 --> 00:11:46,261
Every January, prepare each client's tax situation

177
00:11:46,261 --> 00:11:50,321
and proactively reach out to them before anyone even asks.

178
00:11:50,321 --> 00:11:52,841
This is a new business goal for our business to drive.

179
00:11:53,101 --> 00:11:56,261
And in an AI native firm,

180
00:11:56,481 --> 00:12:00,861
all we have to do is make this goal part of our agent's set of goals.

181
00:12:01,641 --> 00:12:05,161
And because they have the primitives to be able to drive work

182
00:12:05,761 --> 00:12:07,581
through people for the business,

183
00:12:07,841 --> 00:12:09,181
they start driving this goal.

184
00:12:09,561 --> 00:12:11,301
And just like that,

185
00:12:11,601 --> 00:12:15,461
you have these analyses being done in January

186
00:12:15,461 --> 00:12:18,981
and these one-click approval emails

187
00:12:18,981 --> 00:12:23,301
with the full plan for every client showing up on our users' plates.

188
00:12:23,661 --> 00:12:26,461
And because they're aligned with the goal of proactivity,

189
00:12:26,761 --> 00:12:31,881
they can one-click approve and drive the business towards this new outcome.

190
00:12:33,561 --> 00:12:38,661
So this is sort of, I think, the key difference

191
00:12:38,661 --> 00:12:42,901
between a company that uses AI tools

192
00:12:42,901 --> 00:12:47,261
versus a company that is AI native.

193
00:12:47,861 --> 00:12:49,601
in an AI native company

194
00:12:49,601 --> 00:12:50,821
you know you've built one

195
00:12:50,821 --> 00:12:53,141
when you can steer your business

196
00:12:53,141 --> 00:12:55,161
by steering your agents

197
00:12:56,121 --> 00:12:56,681
and

198
00:12:56,681 --> 00:12:59,381
that is our ambition

199
00:12:59,381 --> 00:13:00,261
at Arca

200
00:13:00,261 --> 00:13:03,221
and we are hiring

201
00:13:04,161 --> 00:13:04,721
so

202
00:13:04,721 --> 00:13:06,941
please check out our website

203
00:13:06,941 --> 00:13:09,201
please reach out to me if this is something that you're

204
00:13:09,201 --> 00:13:11,141
interested in working on, these sorts of problems

205
00:13:11,141 --> 00:13:13,261
and thank you for listening

206
00:13:14,401 --> 00:13:14,961
applause