Machine transcription with reviewed terminology and names. Some recognition errors may remain. Timestamps refer to the livestream; subtitle downloads are timed to the session clips.
1:53:23Good 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.
1:53:44and today I'm going to talk to you about ARCA's Long Horizon agents that run our business.
1:53:51ARCA is an AI native services firm.
1:53:55We provide our clients with financial services like financial planning, tax planning, asset management, investment advice,
1:54:04and we serve a range of clients.
1:54:06Some of them are like technology executives that have amassed a lot of wealth and need a real plan to retire early.
1:54:13but also that family in New Hampshire that made their first million dollars
1:54:17and want to buy that house finally.
1:54:22I said we were AI native.
1:54:25There's lots of definitions of this term these days, very overused.
1:54:29What we mean by AI native is that agents run our business.
1:54:36Not that they assist the business from some co-pilot corner.
1:54:40They execute the decisions that shape the outcomes that our business has for our clients.
1:54:48Work in this industry, it's interesting to know what it looks like.
1:54:52In financial services, the work is more long horizon and less long running.
1:54:57And what I mean by that is this is less agents grinding tokens for like hours or days on end, right?
1:55:05This is more one client goal that spans 10 days.
1:55:09It's very stop-start. There's many actors in between, many decision points, many handoffs, and ultimately that goal lands like 10 days later.
1:55:19That is the kind of work that we are working on.
1:55:23And so here's an example. Jane's our client, fairly wealthy, but has a big tax bill.
1:55:28Oh. Has a big tax bill in April 15th, right?
1:55:35So 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.
1:55:45So there's the advisor that will do all of the statement analysis and the analysis of how to make this tax optimal.
1:55:52There's the trader that would actually liquidate the assets.
1:55:57assets themselves will take days to settle, and then the ops people will come in and trigger the ACH.
1:56:03that will itself take days to land.
1:56:06So this is kind of the work I'm talking about in this industry,
1:56:09which takes days to settle.
1:56:12And of course, if you do this at scale for several, several clients,
1:56:16the amount of time our firm takes to do this sort of work exponentiates, right?
1:56:20So at ARCA, we took a different approach.
1:56:25It's helpful to know that ARCA is not a SaaS company.
1:56:29We are not selling software to other businesses, to other firms.
1:56:33We are the firm.
1:56:35We're an applied AI company that happens to be running a services business in this industry.
1:56:40And so the approach we took was, all right, we're going to have agents own this whole long horizon outcome end to end.
1:56:47And for that, we gave them a couple of key components I'll give you a sneak peek of.
1:56:52We gave them signals from the external world that keeps them really aware of what is happening.
1:56:57So that email will be read by agents.
1:57:00We gave them all of our operational context on what to do about that email.
1:57:04What does this industry do with the situation of taxes?
1:57:07We gave them a harness that gives them all of the tools and the guardrails,
1:57:13and more importantly, persistence of memory over the long horizon period
1:57:18to carry that work out.
1:57:20And, of course, we gave them the ability to collaborate with people.
1:57:26So that work with agents compresses a lot,
1:57:31because all we did was we took humans out of the critical path of this work.
1:57:36Nobody has to babysit this goal over days.
1:57:40Agents are on it.
1:57:41In this talk, I'd love to describe to you how we actually do that.
1:57:46Let's follow the journey of this email.
1:57:50Jane owes $100,000 for taxes.
1:57:52this email, much like many other signals in our platform,
1:57:59will be read by background agents.
1:58:01So there's phone calls our firm is making,
1:58:04there's documents that are coming in,
1:58:05there's meetings being had,
1:58:07there's events coming in from our custodian,
1:58:09there's market movement that's happening.
1:58:11These are all signals that our background agents
1:58:13are continuously evaluating to progress work.
1:58:17So this email is one of those signals.
1:58:19And when the signal is ingested by agents,
1:58:21they actually get to the work.
1:58:22right? We have a harness that gives them the knowledge to know what to do about the situation.
1:58:28They have deep client memory, so they know, hey, let's never touch Jane's Apple position,
1:58:34because she said that in a call like eight months ago, right? We, of course, have a sandbox for all
1:58:39of the financial computation that needs to happen for this analysis. But largely, I think harnesses
1:58:45today are coalescing on a standard shape. So it's more interesting what happens outside of
1:58:50this harness.
1:58:54So, after they've done the work, they produce this plan.
1:58:57This plan by itself cannot be executed without people.
1:59:01So, in our product, we build a space, agents build a space, where they put out this plan,
1:59:06and you can chat with the agent to kind of know what's happening with the plan.
1:59:10There's multiple people that can be chatting with the agent.
1:59:14And more importantly, this space is where also they pull people in when they need people
1:59:19to execute some work.
1:59:20So here in this case, we would approve the plan and the email out to Jane.
1:59:26And then there's this whole dance of, has Jane replied?
1:59:30What does her answer mean for the plan that we had?
1:59:34And then all of these other questions that will have to be answered
1:59:37throughout this whole long-horizon work.
1:59:40At Arca, we were like, of course, we're not going to have people babysitting it.
1:59:45So we let our agents watch the world so people don't have to.
1:59:48We give them tools to be able to register interest in anything happening in our world.
1:59:54So they're able to say, let me know when Jane responds.
1:59:58Let me know when that ACH lands.
2:00:00And this is powered by our signal system in the background,
2:00:04where agents are routing signals to agents that are interested in it.
2:00:10And of course, if the signal never lands,
2:00:14agents can wake themselves up on a schedule as well
2:00:17to keep track of work.
2:00:21But then, of course, the real world is very messy, right?
2:00:25People are waiting for that email,
2:00:26but Jay never responds to that email.
2:00:30Say we meet her in person,
2:00:32and we learn some information that affects the plan.
2:00:36So the person goes in,
2:00:37and they add this information into our UI.
2:00:41But this act of adding information or doing work in our UI
2:00:44is also a signal that will wake up an agent.
2:00:48In fact, when people work in our product,
2:00:51when they reschedule a calendar event,
2:00:54that's a signal.
2:00:55When they mark a task complete, that is a signal.
2:00:58So this system lets us be both aware
2:01:02of everything happening in the external world
2:01:04and the internal world as work is progressing
2:01:07so our agents can participate in work.
2:01:12and of course agents are communicating with each other as well
2:01:15so this signal read by a background agent
2:01:17we let it notify another agent about this
2:01:20so that it can update the space to say
2:01:22hey, we just need $80,000 instead of the $100,000
2:01:26so the work has changed because of these signals
2:01:30and the last piece of the story is that
2:01:32agents are able to pull in the right people at the right time
2:01:36So the agent on the left has read operational context
2:01:39to be like, okay, the next step here
2:01:42is to pull in the person that trades these assets.
2:01:46And so that person now gets a task
2:01:48and gets pulled in to participate in this long horizon work.
2:01:53What we've done overall, if you take a step back,
2:01:55is we've given our agents the primitives
2:01:57for driving long horizon work.
2:02:01What do people do at work, right?
2:02:02They need to stay aware of what is happening,
2:02:04so we've built this very intense signal system for our agents to be aware.
2:02:09They need to know what to do next,
2:02:11so we've given them all this operational context.
2:02:13They need to do the work, so it's our harness
2:02:15that lets them actually perform the work and keep track of it.
2:02:18It's long persistence memory systems that hold the state
2:02:24and bring in the right people at the right time.
2:02:29This is a pretty simple idea on the face of it,
2:02:32but there's a lot of emergent complexity.
2:02:35You can imagine at any point,
2:02:37there's hundreds of these signals coming in,
2:02:39and tasks are being added on people's plates,
2:02:44they're being pulled in,
2:02:45background agents are communicating with each other.
2:02:48So over the last six months,
2:02:49we've been working on getting this system to be really good.
2:02:53We've been closing all of our context gaps
2:02:56so that agents are fully aware.
2:02:58We have been working on the precision of tasks,
2:03:01So pulling in the right person at the right time.
2:03:04Work that needs to be tracked for the future is tracked in the agent's memory.
2:03:08Work that is doable now becomes a task, right?
2:03:11So we've been working on our ability to hold state over long horizons as well.
2:03:16And because we are the firm and not just a SaaS company,
2:03:20we've been changing and standardizing how work happens in the
2:03:23firm in the first place.
2:03:25And 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.
2:03:39So I guess there's a question to be asked here, like, why do this? What does this unlock? What is the takeaway here?
2:03:48For us, this is our whole thesis, is that we want our platform to be so AI native that agents drive the business.
2:03:58Agents run the business.
2:04:01They are making the decisions that shape our outcomes.
2:04:04And so at ARCA, when people log in, they see a list of tasks that have been prepared for them by agents.
2:04:12They've already read that email.
2:04:13They've looked at the document that came in.
2:04:16they've done like 90% of the work,
2:04:18and they're pulling you in just when it's time for you to act.
2:04:23So building this system lets us get to a point where at Arca,
2:04:29agents prompt people more than people are prompting agents,
2:04:32which lets our agents drive the work that our firm does.
2:04:37And because agents are driving the work the firm does,
2:04:41agents can drive the business.
2:04:43Let me give you an example.
2:04:44Remember that email that Jane sent us?
2:04:47This is a bit of a failure, right?
2:04:49Because we should have just been proactive about this in the first place, right?
2:04:53So Jane should never have to send this email.
2:04:56So say at, you know, our AI native business, we take on this new goal.
2:05:01Our new goal is proactivity.
2:05:06Every January, prepare each client's tax situation
2:05:09and proactively reach out to them before anyone even asks.
2:05:14This is a new business goal for our business to drive.
2:05:16And in an AI native firm,
2:05:20all we have to do is make this goal part of our agent's set of goals.
2:05:25And because they have the primitives to be able to drive work
2:05:29through people for the business,
2:05:31they start driving this goal.
2:05:33And just like that,
2:05:35you have these analyses being done in January
2:05:39and these one-click approval emails
2:05:42with the full plan for every client showing up on our users' plates.
2:05:47And because they're aligned with the goal of proactivity,
2:05:50they can one-click approve and drive the business towards this new outcome.
2:05:57So this is sort of, I think, the key difference
2:06:02between a company that uses AI tools
2:06:06versus a company that is AI native.
2:06:11in an AI native company
2:06:13you know you've built one
2:06:14when you can steer your business
2:06:16by steering your agents
2:06:19and
2:06:20that is our ambition
2:06:23at Arca
2:06:23and we are hiring
2:06:27so
2:06:28please check out our website
2:06:30please reach out to me if this is something that you're
2:06:32interested in working on, these sorts of problems
2:06:34and thank you for listening
2:06:38applause
