# Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca

Aman Singh, Arca

[Talk page](https://ai.engineer/talks/oP2WLFpMoDY-inside-arcas-long-horizon-agents-that-run) · [Video](https://www.youtube.com/watch?v=oP2WLFpMoDY)

## Inside Arca’s Long-Horizon Agents That Run the Business

Aman Singh follows a client’s tax-payment request through signals, persistent memory, shared plans and human handoffs—and explains how the same machinery could turn a proactive business goal into work.

### One client goal across days of stop-start work

[0:12](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=12s)

Arca provides financial planning, tax planning, asset management and investment advice. Its clients range from technology executives planning an early retirement to a family in New Hampshire that has made its first million dollars and wants to buy a house. Aman Singh, Arca’s AI lead, gives “AI native” a specific operational meaning: agents drive the work that determines client outcomes. Understanding that claim starts with the kind of work a financial services firm has to coordinate.

![Day 1 to Day 12 timeline showing Jane adjusting the plan, the trader raising cash, settlement, the ACH transfer and the advisor confirming with Jane.](https://ai.engineer/api/talk-essay-media/f8dd8fa88f0cbc1606e9f416e83e622daaf4502dfda6f8511d2e03b2feaba21a)

Jane’s example includes two days for settlement and three for ACH transit — recording at 2:41

The useful distinction is **long horizon versus long-running**. A client goal might span ten days without requiring ten days of continuous model inference. Work starts, stops, waits for someone else, resumes after a decision and passes between people. The engineering problem is keeping the goal moving across those interruptions. [1:19](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=79s)

Jane’s tax bill supplies the running example. She needs cash for April 15, and the firm must work out a tax-efficient way to raise it. An advisor analyzes statements and the tax consequences of the proposed sales. A trader liquidates the assets. Those transactions take days to settle; operations staff then initiate an ACH transfer, which itself takes days to arrive. Analysis, execution and waiting all belong to the same client outcome. Repeating this across clients creates a growing coordination burden.

### From Jane’s email to a plan that remembers her preferences

[2:49](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=169s)

Arca operates the services business itself, so it can design both the software and the work agents undertake. Its approach gives agents responsibility for carrying a client outcome across the entire period: external signals keep them aware, operational context tells them how the firm handles a situation, a harness supplies tools and guardrails, persistent memory carries state, and collaboration brings people into the work. Singh describes the benefit as removing the need for someone to “babysit” a goal over days. That reduces human coordination delays; the example’s settlement and transfer waits still remain. [2:49](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=169s)

Jane’s email says she owes **$100,000** in taxes. Background agents read it as a signal that can initiate work. The analysis also draws on client memory: eight months earlier, Jane said in a call that the firm should never touch her Apple position. That preference changes which assets the plan can consider selling. The important relationship is between the new cash requirement and an older constraint—the email alone does not contain everything needed to prepare an appropriate plan.

![Recording frame at 5:11: From Jane’s email to a plan that remembers her preferences](https://ai.engineer/api/talk-essay-media/51799a8fc2b5224b7242f22efea8d43e41c0cb6871b00926b86477abd0af6126)

From Jane’s email to a plan that remembers her preferences — recording at 5:11

The system combines several distinct inputs and capabilities:

- **Signals:** Emails, phone calls, incoming documents, meetings, custodian events and market movements give background agents information that may advance existing work or create new work.
- **Operational knowledge and client memory:** The harness supplies knowledge about handling the situation, while remembered client preferences constrain the proposed action.
- **Financial computation:** A sandbox supports the calculations needed for the analysis.

Together these let an agent move from noticing a request to preparing a plan. The talk identifies tools and guardrails as harness components, but does not specify their enforcement rules or the financial optimization method.

Singh treats the harness as only part of the system. He sees harnesses converging toward a common shape and shifts attention to what happens around them: keeping a plan alive, noticing changes and coordinating the next participant after the initial analysis has finished.

### A shared space for approval, and two ways to wake up

[5:25](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=325s)

Where does the plan become something people can inspect and act on? Agents create a **space** containing the proposed plan. Multiple people can chat with the agent there to understand it, and the agent can pull in someone whose participation is needed. For Jane, a person approves the plan and the outgoing email. This is a consequential limit on end-to-end ownership: the agent coordinates the outcome, while the proposed plan still requires people to execute it. [5:25](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=325s)

![Jane’s $100,000 tax plan and chat appear beside a draft email with Approve and Edit buttons.](https://ai.engineer/api/talk-essay-media/028310a83f2656817e3af71b7baf86fe09382e6c2a24041629c402e5aa48fa5f)

The advisor can review the plan and approve the email to Jane — recording at 5:48

Approval creates a new waiting problem. Has Jane replied? What does her answer mean for the plan? Later, has the ACH arrived? Arca gives agents tools to **register interest** in events such as Jane’s response or the transfer landing. Background agents route signals to the agents interested in them. A person does not have to keep reopening the case to check whether something happened.

An event subscription cannot resolve silence by itself. If the expected signal never arrives, an agent can wake itself on a schedule to keep track of the work. Event notifications and scheduled wake-ups are separate ways to resume attention: one reacts to something happening, while the other gives the agent a chance to revisit work despite no matching event.

In Jane’s case, the information that changes the plan comes through a different channel altogether.

Event notifications and scheduled wake-ups
Signals deliver relevant changes; scheduled wake-ups let the agent revisit work even when an expected signal does not arrive.
- World events — Jane replies, or the ACH transfer lands.
- Signal system — Background agents route signals to interested agents.
- Agent tracking Jane’s work — Registers event interests and can arrange a scheduled wake-up.
- Scheduled wake-up — Provides a separate trigger when the expected signal never arrives.
- Agent tracking Jane’s work → Signal system (Registers interest)
- World events → Signal system (Supplies signals)
- Signal system → Agent tracking Jane’s work (Routes relevant signals)
- Agent tracking Jane’s work → Scheduled wake-up (Schedules a wake-up)
- Scheduled wake-up → Agent tracking Jane’s work (Wakes agent independently of an event)

### Jane never replies, but the plan still changes

[6:52](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=412s)

The next turn breaks the tidy email workflow. Jane never answers the email. Instead, someone meets her in person and learns information that affects the plan. That person enters the information into Arca’s interface, and the act of entering it becomes a signal that wakes an agent. Progress depends on capturing the new information, rather than waiting indefinitely for the original communication channel. [6:52](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=412s)

Internal work generates signals too. Rescheduling a calendar event or marking a task complete can notify the agent system. This extends awareness beyond incoming client communications: people’s actions inside the product can change what an agent should do next.

What changes in Jane’s case? A background agent reads the signal and notifies another agent, which updates the shared space: the cash requirement is now **$80,000 instead of $100,000**. The advisor’s note on the slide says Jane’s bonus arrives March 1 and she wants $20,000 of it applied to the tax bill. It demonstrates how newly captured information reaches the agent maintaining the plan, allowing that plan to change even though the expected email reply never came.

![Recording frame at 7:55: Jane never replies, but the plan still changes](https://ai.engineer/api/talk-essay-media/92e05c91292ff015ecc33cf3958b8b7abe91cd054c9aa8600e9e23bab29e8057)

Jane never replies, but the plan still changes — recording at 7:55

The next handoff comes from operational context. An agent determines that the person who trades these assets needs to participate, creates a task and brings that person into the work. Updated information changes the plan; knowledge of the firm’s process determines who acts next. Those are distinct responsibilities, and the system needs both to carry the goal forward.

### The hard part is deciding what deserves a task

[8:24](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=504s)

The example now contains the main primitives of long-horizon work. Signals maintain awareness. Operational context explains what to do next. The harness performs work, persistent memory holds state across interruptions, and collaboration brings in the appropriate person. These capabilities address different failure points: an agent can know a procedure yet miss a new fact, or notice a fact yet fail to retain the unfinished work it affects.

![Recording frame at 9:56: The hard part is deciding what deserves a task](https://ai.engineer/api/talk-essay-media/1642e3daed6ae7b9832218146c29bf5ab084437f56aa5820d2fb4849973e3ed4)

The hard part is deciding what deserves a task — recording at 9:56

A simple set of primitives produces a complicated operating environment. Singh describes hundreds of signals arriving, background agents communicating and tasks appearing on people’s plates. Over the preceding six months, Arca had worked on closing context gaps, improving task precision and maintaining state over long periods. The difficult question is no longer just whether an agent can analyze a tax request. It is whether the surrounding system brings in the right person at the right time. [9:01](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=541s)

Arca draws a practical distinction between two kinds of unfinished work:

- **Work for later:** Something that needs future tracking stays in the agent’s memory.
- **Work doable now:** Something ready for action becomes a task.

That distinction matters to the human experience. Turning every future concern into an immediate assignment would fill people’s task lists before they could usefully act. Keeping everything in memory would hide work that is ready to move.

Operating the firm gives Arca another way to improve agent behavior: change and standardize how the work happens, then encode that operational context into agent state. The software does not have to accommodate every inconsistent practice as a permanent requirement. Colleagues using the product also supply detailed reasoning that becomes evaluation material, connecting judgments from actual work to improvements in the system.

### People log in to work the agents have already prepared

[10:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=611s)

What does this coordination machinery change for an employee? When people log in, they see tasks agents have prepared after reading the relevant emails and documents. The intended handoff arrives when someone needs to act, with much of the preparation already done. Singh describes agents having completed “ninety percent of the work”; this is his rough characterization, without a defined workload or measurement method. The concrete product behavior is the prepared task list. [10:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=611s)

![Recording frame at 10:51: People log in to work the agents have already prepared](https://ai.engineer/api/talk-essay-media/c83074104830e3561386996f14539950188e6dd9f8a49cb948ff388323970b51)

People log in to work the agents have already prepared — recording at 10:51

This reverses the usual initiation pattern: at Arca, Singh says, agents prompt people more than people prompt agents. A person does not first have to notice an email, ask for an analysis and remember the subsequent handoffs. Agents initiate and coordinate that preparation, then request human participation. Control over which work gets initiated and advanced is what connects the agent platform to the direction of the business.

### The final correction: Jane should never have needed to ask

[11:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=671s)

The ending revisits the original trigger with a sharper judgment: Jane’s email is itself “a bit of a failure.” A firm aiming to be proactive should have prepared for her tax situation before she asked. Successfully carrying a reactive request through its handoffs still leaves the business dependent on the client noticing the need first. [11:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=671s)

![Recording frame at 12:14: The final correction: Jane should never have needed to ask](https://ai.engineer/api/talk-essay-media/2636a5255f5a31b02948bc781a5148cf6c329f7e22637a777d2743751baf9821)

The final correction: Jane should never have needed to ask — recording at 12:14

Singh proposes a new business goal: **every January, prepare each client’s tax situation and reach out proactively**. Adding that goal to the agents’ goals would give the existing machinery a different reason to start work. The same capabilities—client context, analysis, persistent state and human coordination—would support January preparation without waiting for individual requests.

The proposed visible result is a full plan and an email ready for one-click approval on each user’s task list. People would approve the prepared outreach, moving the firm toward the new goal. This is presented as an illustration of how the architecture could support a change in business direction, rather than a measured rollout of the January program. The meaningful change is from a client email initiating one case to a recurring business objective initiating preparation across clients.

That is Arca’s ambition for an AI-native firm: “steer your business by steering your agents.” It depends on everything built earlier in the example. Goals need agents that can notice relevant events, preserve constraints, maintain unfinished work and bring people in when action is possible. The closing invitation to work on these problems follows from that operating model: the project spans both agent infrastructure and the design of the services business itself.

## Key takeaways

- Long-horizon work preserves one goal across days of decisions, handoffs and waiting. Continuous model execution is only a small part of that problem.
- A useful plan combines new signals with older constraints: Jane’s $100,000 tax request must respect her instruction from eight months earlier to preserve her Apple position.
- Event subscriptions resume work when something happens; scheduled wake-ups let agents revisit work when the expected event never arrives.
- Internal product actions also generate signals. In Jane’s example, information from an in-person meeting reaches agents and changes the cash requirement to $80,000.
- Keep future work in agent memory and turn actionable work into tasks. Task precision determines whether human participation happens at a useful moment.
- The closing January tax-planning example extends the architecture from responding to client requests toward initiating work from a business goal, with people approving the prepared plans and outreach.

## Related talks

- [How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase](https://ai.engineer/talks/_4Ph2dhmUFI-know-when-ai-agent-work-unsupervised-evan) — Jane’s advisor approves the outgoing email, a concrete point of comparison for the question of when an agent can work unsupervised.
- [Is the Agent Harness Dead? — Jediah Katz, SpaceXAI](https://ai.engineer/talks/7Y4sbZgBdnc-is-agent-harness-dead-jediah-katz-spacexai) — Singh’s distinction between what the harness does and what happens around it gives context for Katz’s question about the future of the agent harness.
- [Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse](https://ai.engineer/talks/obfPzbapMIM-evals-high-stakes-ai-agents-doneyli-de) — Arca uses colleagues’ reasoning to build evaluations, connecting Singh’s example to the broader topic of evaluating high-stakes agents.

## Slides

### What AI-native means at Arca

[1:03](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=63s) [The opening slide says that agents run the business.](https://assets.ai.engineer/sites/aiecode2025/sha256/c3c48b7b88ff7c18e7ef716512d393bae7659772261a731b198f9142cff10532.webp)

What do we mean by "AI native"?
Agents run the business

### Long-horizon work

[1:23](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=83s) [The slide distinguishes work that spans a long period from a process running continuously.](https://assets.ai.engineer/sites/aiecode2025/sha256/e54c6beb90fdbe3db93a49c8c19d19cc9350b0d1ddca76be5b93c9bb655bd632.webp)

What work looks like in this industry
Long-horizon, not long-running

### Jane's April tax payment

[1:52](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=112s) [A checklist breaks a hypothetical $100,000 payment into steps from choosing accounts through confirming receipt.](https://assets.ai.engineer/sites/aiecode2025/sha256/800c07a3e2de344c478e2a997a8d05c24058378c36679a40b2fd65b745fd8d0b.webp)

Jane owes $100K by April 15
I Work out which accounts to draw from
3 Make it tax-efficient
3 Get Jane's approval
Raise the cash
5)
Let the sales settle
Send the ACH and wait for it to land
7 Confirm with Jane
ARCA

### A payment across time

[2:10](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=130s) [A timeline shows the dependencies and waiting periods in the example.](https://assets.ai.engineer/sites/aiecode2025/sha256/7ca9e20521d14bbff36a978166918b1c0af9e012fe93e4e558d04d67255d6a72.webp)

Long-horizon work
Jane and
the world
Adjusts the plan
Jane
1 day
Settling
2 days
The firm
Advisor
1 day
Plans the sale
Trader
1 day
Raises cash
Day 1
Day 2
Day 3
Day 4
ARCA
Ops
2 days
Verifies, sends the ACH
Day 6
Day 8
ACH in transit
3 days
Advisor
1 day
Confirms with Jane
Day 11 Day 12
6

### An applied AI services business

[2:59](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=179s) [Arca describes its operating model as a company delivering services with applied AI.](https://assets.ai.engineer/sites/aiecode2025/sha256/de1df082b7eecd6c7b4556721602ced5451a6097e5f9642b964d201ef8897782.webp)

Arca is not a SaaS company
We're an applied AI company
running a services business

### Owning an outcome end to end

[3:51](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=231s) [Four columns identify signals, operational context, a harness and people.](https://assets.ai.engineer/sites/aiecode2025/sha256/563dd83959452ba7400a8d14b5a2b6c32c4cb3a88b9c628d6fa276a09d94ded4.webp)

Agents own outcomes end to end
Signals
Every email, call, document and
event, from inside the firm and
out.
Operational context
The firm's knowledge and each
client's history, to decide what to
do.
Harness
Tools to act with guardrails,
memory of the work in flight, and
the persistence to hold it for
days.
ARCA
People
Pulled in at the right moment,
with everything they need to
decide.

### Questions that persist between steps

[6:05](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=365s) [The checklist asks about a client reply, trade settlement, the transfer and blocked work.](https://assets.ai.engineer/sites/aiecode2025/sha256/5c95ef593409dee00bbbbf97ee8742def02be8f00c6b0d13f1e72014ad5a095a.webp)

Has Jane replied?
What does her answer change?
Did the trades settle?
Did the ACH land?
Is anything stuck?
ARCA
14

### Watching for the next signal

[6:41](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=401s) [Example conditions show what should wake an agent and when to follow up.](https://assets.ai.engineer/sites/aiecode2025/sha256/9923237c1bda648d8355455594ed1e34f1fdd00150203aed1d9eaa874fa67be1.webp)

Agents watch the world, so people don't have to
Wake on an event
notify_me (when="Jane responds to this email thread")
notify_me (when= "Peter approves the trade and it's issued to the custodian")
notify_me (when= "The ACH lands")
notify_me (when= "Jane uploads her tax return")
Wake at a time
remind_me (when= "in 2 days")
remind_me (when= "Friday at 9am")
remind_me (when="a week before April 15")
ARCA
15

### Primitives for long-horizon work

[8:58](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=538s) [A two-column table pairs the work people do with four components of an agent system.](https://assets.ai.engineer/sites/aiecode2025/sha256/b9a0a5d3af6324769c3a328e74e6df546c69792ba9366076c7ca2307f9eeb8ff.webp)

Primitives for long-horizon work
• People
Stay aware of what's happening
Know how the firm works
Do the work, and keep track of it
Bring in the right person
• Arca agents
Signals
Every email, call, edit and event
Operational context
The firm's knowledge and each client's history
Harness
Tools, memory, and the persistence to hold work for days
People
Tasks for the right person, at the right time
ARCA
21

### Preparing cash before a request

[11:31](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=691s) [A business goal asks agents to prepare each client's cash for tax season before anyone asks.](https://assets.ai.engineer/sites/aiecode2025/sha256/30ff6797ac2811bc19ed5487d94fbea051d08748110b0281c48167f15f0b8f5f.webp)

A new business goal: proactivity
Every January, prepare each client for tax season before anyone asks.
ARCA
26

### Steering agents

[12:32](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=752s) [The closing slide connects steering the business with steering its agents.](https://assets.ai.engineer/sites/aiecode2025/sha256/198943290cbd9867916468fdc2330ab89faf556a86b13a8e574961118a439646.webp)

Steer your business by
steering your agents

## Full transcript

[0:12](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=12s) My name's Aman. Uh, I lead AI at Arca. Um, and today I'm gonna talk to you about Arca's long-horizon agents that run our business. Arca is an AI-native services firm. We provide our clients with financial services like financial planning, tax planning, asset management, investment advice, um, and we serve a range of clients. Some of them are like technology executives that, um, have amassed a lot of wealth and need a real

[0:42](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=42s) plan to retire early. But also that family in New Hampshire that made their first million dollars and, and want to buy that house finally. Um, I said we were AI native. There's a lots of, there's lots of definitions of this, this term these days, very overused. But what we mean by AI native is that agents run our business. Not that they assist the business from some co-pilot corner. They, they

[1:12](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=72s) execute the decisions that shape the outcomes that our business has for our clients.

[1:19](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=79s) Work in this industry, it's interesting to know what it looks like. In financial services, the work is more long horizon and less long-running. And what I mean by that is this is less agents grinding tokens for like hours or, or days on end, right? This is more one client goal that spans ten days. It's very stop-start. There's many actors in between, many decision points, many handoffs, um, and ultimately that goal lands

[1:49](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=109s) like ten days later. That is the kind of work that, that we are working on. And so here's an example. Jane's our client, fairly wealthy, but has a big tax bill. Oh. Um, has a big tax bill in, um, uh, in, in April fifteenth, right? So we're gonna help Jane figure out what is the most tax optimal way for this cash to be, to be raised to pay the tax bill. So there's the advisor that will do all of the, the

[2:19](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=139s) statement analysis and the analysis of, of how to make this tax optimal. There's the, the trader that would actually liquidate the, the assets. 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. That will itself take days, uh, to land. So this is kind of the work I'm, I'm talking about in this industry, which takes da-days to settle. And of course, if you do this at scale for several, several clients, the amount of time our firm takes to do this

[2:49](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=169s) sort of work exponentiates, right? So at Arca, we took a different approach. It's helpful to know that Arca is not a SaaS company. We are not selling software to other businesses, to other firms. We are the firm. We're an applied AI company that happens to be running a services business in this industry. And so the approach we took was, all right, we're gonna have agents own this whole long-horizon outcome end to end. And

[3:19](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=199s) for that, we gave them f- couple of key components I'll give you a sneak peek of. We gave them signals from the external world that keeps them really aware of what is happening. So that email will be read by agents. We give them all of our operational context on what to do about that email. What does this industry do with this situation of taxes? Right? We gave them a harness that gives them all of the tools and the guardrails, and more importantly, persistence of memory over the, the kind of long horizon

[3:48](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=228s) period to carry that work out. And of course, we gave them the ability to collaborate with people.

[3:57](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=237s) 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. Nobody has to babysit this, this, uh, this goal over days. Agents are, are on it. In this talk, I'd love to describe to you how we actually do that.

[4:17](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=257s) Let's follow the journey of this email.

[4:21](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=261s) Jane owes a hund- hundred thousand dollars for taxes. This email, much like many other signals in our, in our platform, will be, will be read by background agents. 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. These are all signals that our background agents are continuously evaluating to progress work. So this email is one of those signals. And when the signal is,

[4:51](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=291s) is ingested by agents, they actually get to the work, right? Um, we have a harness that gives them the knowledge to know what to do about the situation. They have deep client memory, so they know, "Hey, let's never touch Jane's Apple position," 'cause, uh, she said that in a call like eight months ago, right? Um, we of course have a sandbox for all of the financial computation that needs to happen for this analysis. Um, but largely, I think harnesses today are, are coalescing on a standard shape. So, um, it's more interesting what happens outside

[5:21](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=321s) of, outside of this harness.

[5:25](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=325s) So after they've done the work, they produce this plan. This plan by itself cannot be executed without people. 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. There's mult-multiple people that can be chatting with the agent. And more importantly, this space is where also they pull people in when they need people to execute some work. So here in this case, we would approve, uh, the plan and the email out to

[5:55](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=355s) Jane. And then there's this whole dance of has Jane replied? What does her, what does her answer mean for the plan that we have, right? And then all of these other, these other questions that will have to be answered throughout this whole, like, long-horizon work. At Arca, we were like, "Of course we're not gonna have people babysitting it." So we let our agents watch the world so people don't have to, right? So we give them tools to be able to register interest in anything happening in our world, so

[6:25](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=385s) they're able, able to say, "Let me know when Jane responds. Let me know when that ACH lands," right? And this is powered by our signal system in, in the background, where agents are routing, uh, signals to agents that are interested in it. And of course, if the signal never lands, agents can, like, wake themselves up on, on a schedule as, as well to keep track of work.

[6:52](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=412s) But then, of course, the real world is very messy, right? People are waiting for that email, but Jane never responds to that email. Say we, we meet her in person and we learn some information that affects the plan. So the person goes in and they, they add this information into our UI. But this act of adding information or doing work in our UI is also a signal that will wake up an agent. In fact, when people work in our pr- in our

[7:22](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=442s) product, when they reschedule a calendar event, you know, that's a signal. When they mark a task complete, that is a signal. 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.

[7:43](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=463s) And of course, agents are communicating with each other as well. So this signal read by a background agent, we let it notify another agent about this so that it can update the space to say, "Hey, we just need, uh, need eighty thousand dollars instead of the hundred thousand." So the work has changed because of, because of these signals.

[8:02](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=482s) And the last piece of the story is that agents are able to pull in the right people at the right time. So, um, 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 these, uh, these assets." And so that person now gets a task and gets pulled in to participate in this long-horizon work.

[8:24](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=504s) What we've done overall, if you take a step back, is we've given our agents the primitives for driving long-horizon work. What do people do at work, right? 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. They need to know what to do next, so we've given them all of this operational context. They need to do the work, so it's our harness that lets them actually perform the work and keep track of it, right? It's long persistence, um, memory systems that, that hold the state

[8:54](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=534s) and bringing the right people at the right time.

[9:01](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=541s) This is a pretty simple idea on the face of it, but there's a lot of emergent complexity. You can imagine at any point there's 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, right? So over the last six months we've, we've been working on getting this system to be really good. We've been closing all of our context gaps so that agents are fully aware. We have been working on the

[9:30](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=570s) precision of tasks, so pulling in the right person at the right time. Work that needs to be tracked for the future is tracked in the agent's memory. Work that is doable now becomes a task, right? So we've been working on, uh, on our ability to hold state over long horizons as well. 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, um, our agent's state.

[10:01](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=601s) And because these are our colleagues that use our product, we collect rich reasoning from them that become evals for, for our system to improve.

[10:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=611s) So I guess there's, there's a question to be asked here, like, "Why do this," right? What does this unlock? What is the takeaway here, right? For, for us, this is our whole thesis, is that we want our platform to be so AI-native that agents drive the business. Agents run the business. They, they are making the decisions that shape our outcomes. And so at Arca, when people log in, right, they see a list of tasks that have been prepared for

[10:41](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=641s) them by agents. They've already read that email. They've, like, looked at the document that came in. They've done, like, ninety percent of the work, um, and they're pulling you in just when it's time for you to act. 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 does. And because agents are driving the work the

[11:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=671s) firm does, agents can drive the business. Let me give you an example. Remember that email that Jane sent us? This is a bit of a failure, right, because we should've just been proactive about this in the first place, right? So Jane should never have to send this email. So say at, you know, our AI-native business, we take on this new goal. Our new goal is proactivity. Every January, prepare each client's tax

[11:40](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=700s) situation and proactively reach out to them before anyone even asks. This is a new business goal for our business to drive, and in an AI-native firm, all we have to do is make this goal part of our agent's set of goals. 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

[12:10](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=730s) and these, these one-click approval emails with the full plan for every client showing up on our users' plates. And because they're aligned with the goal of proactivity, they can one-click approve and drive the business towards this, this new outcome.

[12:28](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=748s) 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. In an AI-native company, you know you've built one when you can steer your business by steering your agents, and that is our ambition at, at Arca, and we are, uh, we're hiring.

[12:58](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=778s) Um, so please, uh, please check out our website. Please reach out to me if this is something that you- you're, you're interested in working on these sorts of problems, and, um, thank you for listening.
