WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=oP2WLFpMoDY

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/oP2WLFpMoDY.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.760 --> 00:00:14.040
My name's Aman.

s3
00:00:14.040 --> 00:00:16.120
Uh I lead AI at Arca.

s4
00:00:16.120 --> 00:00:22.920
Um and today I'm going to talk to you about Arca's long horizon agents that run our business.

s5
00:00:22.960 --> 00:00:26.680
Arca is an AI native services firm.

s6
00:00:26.680 --> 00:00:35.200
We provide our clients with financial services like financial planning, tax planning, asset management, investment advice.

s7
00:00:35.200 --> 00:00:37.160
Um and we serve a range of clients.

s8
00:00:37.160 --> 00:00:44.520
Some of them are like technology executives that um have amassed a lot of wealth and need a real plan to retire early.

s9
00:00:44.520 --> 00:00:52.360
But also that family in New Hampshire that made their first million dollars and and want to buy that house finally.

s10
00:00:52.360 --> 00:00:56.200
Um I said we were AI native.

s11
00:00:56.200 --> 00:01:00.960
There's a lots of There's lots of definitions of this this term these days, very overused.

s12
00:01:00.960 --> 00:01:07.440
But what we mean by AI native is that agents run our business.

s13
00:01:07.440 --> 00:01:11.280
Not that they assist the business from some co-pilot corner.

s14
00:01:11.280 --> 00:01:18.760
They They execute the decisions that shape the outcomes that our business has for our clients.

s15
00:01:19.560 --> 00:01:23.120
Work in this industry, it's interesting to know what it looks like.

s16
00:01:23.120 --> 00:01:28.720
In financial services, the work is more long horizon and less long running.

s17
00:01:28.720 --> 00:01:36.240
And what I mean by that is this is less agents grinding tokens for like hours or days on end, right?

s18
00:01:36.240 --> 00:01:40.840
This is more one client goal that spans 10 days.

s19
00:01:40.840 --> 00:01:42.000
It's very stop-start.

s20
00:01:42.000 --> 00:01:47.560
There's many actors in between, many decision points, many hand-offs.

s21
00:01:47.560 --> 00:01:50.360
Um and ultimately that goal lands like 10 days later.

s22
00:01:50.360 --> 00:01:54.400
That is the kind of work that that we're working on.

s23
00:01:54.400 --> 00:01:55.840
And so here's an example.

s24
00:01:55.840 --> 00:01:59.880
Jane's our client, fairly wealthy, but has a big tax bill.

s25
00:01:59.880 --> 00:02:01.840
Oh.

s26
00:02:01.840 --> 00:02:06.520
Um has a big tax bill in uh in in April 15th, right?

s27
00:02:06.520 --> 00:02:16.480
So, we're going to help Jane figure out what is the most tax optimal way for this cash to be to be raised to pay the tax bill.

s28
00:02:16.480 --> 00:02:23.560
So, there's the advisor that will do all of the the statement analysis and the analysis of of how to make this tax optimal.

s29
00:02:23.560 --> 00:02:28.600
There's the the trader that would actually liquidate the the assets.

s30
00:02:28.600 --> 00:02:34.600
They the assets themselves will take like days to settle, and then the ops people will come in and kind of like trigger the ACH.

s31
00:02:34.600 --> 00:02:37.440
That will itself take days uh to land.

s32
00:02:37.440 --> 00:02:44.040
So, this is kind of the work I'm I'm talking about in this industry, which takes like days to settle.

s33
00:02:44.040 --> 00:02:51.800
And of course, if you do this at scale for several several clients, the amount of time our firm takes to do this sort of work exponentiates, right?

s34
00:02:51.800 --> 00:02:56.160
So, at Arca, we took a different approach.

s35
00:02:56.160 --> 00:03:00.560
It's helpful to know that Arca is not a SaaS company.

s36
00:03:00.560 --> 00:03:04.360
We are not selling software to other businesses, to other firms.

s37
00:03:04.360 --> 00:03:06.400
We are the firm.

s38
00:03:06.400 --> 00:03:11.360
We're an applied AI company that happens to be running a services business in this industry.

s39
00:03:11.360 --> 00:03:18.800
And so, the approach we took was, all right, we're going to have agents own this whole long horizon outcome end to end.

s40
00:03:18.800 --> 00:03:23.560
And for that, we gave them a couple of key components, I'll give you a sneak peek of.

s41
00:03:23.560 --> 00:03:28.840
We gave them signals from the external world that keeps them really aware of what is happening.

s42
00:03:28.840 --> 00:03:31.560
So, that email will be read by agents.

s43
00:03:31.560 --> 00:03:35.120
We give them all of our operational context on what to do about that email.

s44
00:03:35.120 --> 00:03:38.519
What does this industry do with this situation of taxes?

s45
00:03:38.519 --> 00:03:38.760
Right?

s46
00:03:38.760 --> 00:03:49.519
We gave them a harness that gives them all of the tools and uh the guardrails, and more importantly, persistence of memory over the the kind of long horizon period

s47
00:03:49.519 --> 00:03:51.440
to carry that work out.

s48
00:03:51.440 --> 00:03:56.560
And of course, we gave them the ability to collaborate with people.

s49
00:03:57.320 --> 00:04:06.880
So, that work with um with agents compresses a lot because all we did was we took humans out of the critical path of of this work.

s50
00:04:06.880 --> 00:04:11.240
Nobody has to babysit this this uh this goal over days.

s51
00:04:11.240 --> 00:04:12.480
Agents are are on it.

s52
00:04:12.480 --> 00:04:16.880
In this talk, I'd love to describe to you how we actually do that.

s53
00:04:17.880 --> 00:04:20.920
Let's follow the journey of this email.

s54
00:04:21.440 --> 00:04:24.360
Jane owes on the $100,000 for taxes.

s55
00:04:24.360 --> 00:04:33.000
This email, much like many other signals in our in our platform, will be will be read by background agents.

s56
00:04:33.000 --> 00:04:42.040
So, there's like phone calls our firm is making, there's documents that are coming in, there's meetings being had, there's events coming in from our custodian, there's market movement that's happening.

s57
00:04:42.040 --> 00:04:48.440
These are all signals that our background agents are continuously evaluating to progress work.

s58
00:04:48.440 --> 00:04:50.600
So, this email is one of those signals.

s59
00:04:50.600 --> 00:04:54.800
And when the signal is is by agents, they actually get to the work, right?

s60
00:04:54.800 --> 00:04:59.840
Um we have a harness that gives them the knowledge to know what to do about the situation.

s61
00:04:59.840 --> 00:05:08.400
They have deep client memory, so they know, "Hey, let's never touch Jane's Apple position cuz uh she said that in a call like 8 months ago, right?"

s62
00:05:08.400 --> 00:05:14.200
Um we of course have a sandbox for all of the financial computation that needs to happen for this analysis.

s63
00:05:14.200 --> 00:05:18.880
Um but largely, I think harnesses today are are coalescing on a standard shape.

s64
00:05:18.880 --> 00:05:24.400
So, um it's more interesting what happens outside of outside of this harness.

s65
00:05:25.640 --> 00:05:28.919
So, after they've done the work, they produce this plan.

s66
00:05:28.919 --> 00:05:32.240
This plan by itself cannot be executed without people.

s67
00:05:32.240 --> 00:05:41.720
So, in our product, we build a space, agents build a space, where they put out this plan, and there's you can chat with the agent to like kind of know what's happening with the plan.

s68
00:05:41.720 --> 00:05:45.520
There's mult- multiple people that can be chatting with the agent.

s69
00:05:45.520 --> 00:05:51.800
And more importantly, this space is where also they pull people in when they need people to execute some work.

s70
00:05:51.800 --> 00:05:57.640
So here in this case, we would approve uh the plan and the email out to Jane.

s71
00:05:57.640 --> 00:06:01.680
And then there's this whole dance of has Jane replied?

s72
00:06:01.680 --> 00:06:05.560
What does her What does her answer mean for the plan that we had, right?

s73
00:06:05.560 --> 00:06:11.600
And then all of the these other questions that will have to be answered throughout this whole like long horizon work.

s74
00:06:11.600 --> 00:06:15.920
At Arca, we were like, of course we're not going to have people babysitting it.

s75
00:06:15.920 --> 00:06:18.720
So we let our agents watch the world.

s76
00:06:18.720 --> 00:06:19.840
So people don't have to, right?

s77
00:06:19.840 --> 00:06:25.560
So we give them tools to be able to register interest in anything happening in our world.

s78
00:06:25.560 --> 00:06:29.320
So they're able to say, let me know when Jane responds.

s79
00:06:29.320 --> 00:06:31.720
Let me know when that ACH lands.

s80
00:06:31.720 --> 00:06:31.960
Right?

s81
00:06:31.960 --> 00:06:41.880
And this is powered by our signal system in the in the background where agents are routing uh signals to agents that are interested in it.

s82
00:06:41.880 --> 00:06:51.400
And of course, if the signal never lands, agents can like wake themselves up on on a schedule as well to keep track of work.

s83
00:06:52.280 --> 00:06:56.440
But then, of course, the real world is very messy, right?

s84
00:06:56.440 --> 00:07:01.880
People are waiting for that email, but Jane never responds to that email.

s85
00:07:01.880 --> 00:07:07.360
Say we we meet her in person and we learn some information that affects the plan.

s86
00:07:07.360 --> 00:07:11.720
So the person goes in and they they add this information into our UI.

s87
00:07:11.720 --> 00:07:19.600
But this act of adding information or doing work in our UI is also a signal that will wake up an agent.

s88
00:07:19.600 --> 00:07:26.800
In fact, when people work in our in our product, when they reschedule a calendar event, you know, that's a signal.

s89
00:07:26.800 --> 00:07:29.960
When they mark a task complete, that is a signal.

s90
00:07:29.960 --> 00:07:41.360
So this system lets us be both aware of everything happening in the external world and the internal world as work is progressing so our agents can participate in work.

s91
00:07:43.240 --> 00:07:46.560
And of course, agents are communicating with each other as well.

s92
00:07:46.560 --> 00:07:56.000
So, this signal read by a background agent will let it notify another agent about this so that it can update the space to say, "Hey, we just need a need $80,000

s93
00:07:56.000 --> 00:07:57.440
instead of the $100,000."

s94
00:07:57.440 --> 00:08:01.400
So, the work has changed because of because of these signals.

s95
00:08:02.080 --> 00:08:06.960
And the last piece of the story is that agents are able to pull in the right people at the right time.

s96
00:08:06.960 --> 00:08:15.360
So, uh the agent on the left has read our operational context to be like, "Okay, this the next step here is to pull in the the person that trades

s97
00:08:15.360 --> 00:08:17.160
these uh these assets."

s98
00:08:17.160 --> 00:08:24.000
And so, that person now gets a task and gets pulled in to participate in this long-horizon work.

s99
00:08:24.560 --> 00:08:32.080
What we've done overall, if you take a step back, is we've given our agents the primitives for driving long-horizon work.

s100
00:08:32.080 --> 00:08:33.400
What do people do at work, right?

s101
00:08:33.400 --> 00:08:40.200
They They need to stay aware of what is happening, so we've built this very intense like signal system for our agents to be aware.

s102
00:08:40.200 --> 00:08:44.680
They need to know what to do next, so we've given them all of this operational context.

s103
00:08:44.680 --> 00:08:49.520
They need to do the work, so it's our harness that lets them actually perform the work and keep track of it, right?

s104
00:08:49.520 --> 00:08:54.600
It's long-persistence um memory systems that that hold the state.

s105
00:08:54.600 --> 00:08:58.720
And bringing in the right people at the right time.

s106
00:09:00.800 --> 00:09:06.800
This is a pretty simple idea on the face of it, but there's a lot of emergent complexity.

s107
00:09:06.800 --> 00:09:18.520
You can imagine at any point there is hundreds of these signals coming in, and they are, you know, tasks tasks are being added on people's plates, they're being pulled in, background agents are communicating with each other,

s108
00:09:18.520 --> 00:09:19.560
right?

s109
00:09:19.560 --> 00:09:24.480
So, over the last 6 months, we've we've been working on getting this system to be really good.

s110
00:09:24.480 --> 00:09:29.640
We've been closing all of our context gaps so that agents are fully aware.

s111
00:09:29.640 --> 00:09:35.360
We have been working on the precision of tasks, so pulling in the right person at the right time.

s112
00:09:35.360 --> 00:09:39.720
Work that needs to be tracked for the future is tracked in the agent's memory.

s113
00:09:39.720 --> 00:09:42.640
Work that is doable now becomes a task, right?

s114
00:09:42.640 --> 00:09:47.520
So, we've been working on on our ability to hold state over long horizons as well.

s115
00:09:47.520 --> 00:09:59.160
And because we are the firm and not just a SaaS company, we've been changing and standardizing how work happens in the firm in the first place and operationalizing that context into

s116
00:09:59.160 --> 00:10:00.920
our agent's state.

s117
00:10:00.920 --> 00:10:09.920
And because these are our colleagues that use our product, we collect rich reasonings from them that become evals for for our system to improve.

s118
00:10:11.000 --> 00:10:15.720
So, I guess there's there's a question to be asked here, like why do this, right?

s119
00:10:15.720 --> 00:10:17.920
What does this unlock?

s120
00:10:17.920 --> 00:10:19.840
What is the takeaway here, right?

s121
00:10:19.840 --> 00:10:23.480
For for us, this is our whole thesis.

s122
00:10:23.480 --> 00:10:29.880
Is that we want our platform to be so AI native that agents drive the business.

s123
00:10:29.880 --> 00:10:31.720
Agents run the business.

s124
00:10:31.720 --> 00:10:35.360
They they are making the decisions that shape our outcomes.

s125
00:10:35.360 --> 00:10:43.560
And so, at Arca, when people log in, right, they see a list of tasks that have been prepared for them by agents.

s126
00:10:43.560 --> 00:10:44.960
They've already read that email.

s127
00:10:44.960 --> 00:10:47.400
They've like looked at the document that came in.

s128
00:10:47.400 --> 00:10:54.720
They've done like 90% of the work and they're pulling you in just when it's time for you to act.

s129
00:10:54.720 --> 00:11:07.920
So, building this system lets us get to a point where at Arca, agents prompt people more than people are prompting agents, which lets our agents drive the work that our our firm

s130
00:11:07.920 --> 00:11:09.120
does.

s131
00:11:09.120 --> 00:11:15.080
And because agents are driving the work the firm does, agents can drive the business.

s132
00:11:15.080 --> 00:11:16.120
Let me give you an example.

s133
00:11:16.120 --> 00:11:18.600
Remember that email that Jane sent us?

s134
00:11:18.600 --> 00:11:20.000
This is a bit of a failure, right?

s135
00:11:20.000 --> 00:11:24.880
Because we should have just been proactive about this in the first place, right?

s136
00:11:24.880 --> 00:11:27.280
So, Jane should never have to send this email.

s137
00:11:27.280 --> 00:11:32.800
So, say at you know, our AI native business, we take on this new goal.

s138
00:11:32.800 --> 00:11:36.920
Our new goal is proactivity.

s139
00:11:37.440 --> 00:11:45.320
Every January, prepare each client's tax situation and proactively reach out to them before anyone even asks.

s140
00:11:45.320 --> 00:11:47.880
This is a new business goal for our business to drive.

s141
00:11:47.880 --> 00:11:56.600
And in an AI native firm, all we have to do is make this goal part of our agents' set of goals.

s142
00:11:56.600 --> 00:12:10.360
And because they have the primitives to be able to drive work through people for the business, they start driving this goal, and just like that, you have you have these analyses being done in January,

s143
00:12:10.360 --> 00:12:18.720
and these these one-click approval emails with the full plan for every client showing up on our users' plates.

s144
00:12:18.720 --> 00:12:28.080
And because they're aligned with the goal of proactivity, they can one-click approve and drive the business towards this this new outcome.

s145
00:12:28.680 --> 00:12:42.880
So, this is sort of I think the key difference between a a company that uses AI tools versus a a company that is AI native.

s146
00:12:42.880 --> 00:12:50.880
In an AI native company, you know you've built one when you can steer your business by steering your agents.

s147
00:12:50.880 --> 00:12:58.560
And that is our ambition at at Arca, and we are we're hiring.

s148
00:12:58.560 --> 00:13:01.600
Um so, please uh please check out our website.

s149
00:13:01.600 --> 00:13:06.320
Please reach out to me if this is something that you you're you're interested in working on these sorts of problems.

s150
00:13:06.320 --> 00:13:09.284
And um thank you for listening.

s151
00:13:09.284 --> 00:13:11.284
[applause]