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

Aman Singh, Arca | AI Engineer NYC 2026 | 13:16

Source: https://www.youtube.com/watch?v=oP2WLFpMoDY
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/inside-arcas-long-horizon-agents-that-run-the-business
Published: 2026-10-23
Tags: agents, human-in-the-loop, memory, workflows

## TL;DR
- Arca uses agents to own client goals that unfold across days of stop-start work, from the initial signal through human approvals and settlement.
- A signal system lets agents react to emails, calls, documents, custodian events, market movement, and actions inside the firm's product.
- Arca aims to steer the business by changing the goals given to agents, such as making agents prepare every client's tax situation proactively each January.

## Summary
Aman Singh describes Arca as an AI-native financial services firm where agents run operational work rather than merely assist employees. He contrasts long-horizon work with long-running model execution: a client goal may span ten days, with pauses, hand-offs, decisions, and settlement delays. His example is raising cash for a client's tax bill. Agents read incoming signals, use operational context and financial tools, remember client preferences, produce a plan, and bring people into a shared space when approval or execution is needed. Arca's signal system also watches for replies, ACH settlement, calendar changes, completed tasks, and other events. Singh says the firm has spent the last six months closing context gaps, improving task precision, preserving state, and standardising work so it can be encoded for agents. The intended outcome is that agents prepare tasks and prompt employees when action is needed. Longer term, Arca wants business goals such as proactive tax planning to be implemented by steering the agents that drive the work.

## Key ideas
### Long-horizon work pauses between decisions and settlement
[01:19](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=79s)
Singh distinguishes long-horizon work from agents running continuously for hours or days. In financial services, one client goal can span ten days while work stops and starts around different actors, decisions, hand-offs, and settlement periods. He gives the example of raising cash for Jane's tax bill. An advisor analyses the statements and tax impact, a trader liquidates assets, the assets take days to settle, and operations triggers an ACH transfer that also takes days to arrive. Repeating this process for many clients creates a large operational burden.

### Agents own the client's outcome across the whole process
[02:51](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=171s)
Arca gives agents responsibility for the full long-horizon outcome. They receive signals from the external world, read the firm's operational context, and use a harness with tools and guardrails. Persistent memory lets them carry work across the period between events. Agents can also collaborate with people when the work requires human action. Singh says this removes people from the critical path of watching a goal over several days, while people still approve plans and perform the actions that agents cannot execute alone.

### A signal system keeps agents aware of work as events happen
[04:17](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=257s)
The email about Jane's tax bill is one signal among many. Background agents evaluate phone calls, incoming documents, meetings, custodian events, and market movement. Agents can register interest in future events, such as Jane's reply or the arrival of an ACH. The signal system routes those events to interested agents. If an expected signal never arrives, an agent can wake itself on a schedule. This lets the system continue work without a person repeatedly checking whether something has happened.

### Shared spaces bring people into the process when their action is needed
[05:28](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=328s)
After doing its analysis, an agent produces a plan in a shared space. People can chat with the agent there, and the agent can pull in the person needed for a particular step. For Jane's case, someone might approve the plan and the email, then the system waits for her response and interprets what that response means for the plan. The agent also uses operational context to identify the person who trades the assets, then gives that person a task at the right point in the process.

### Actions inside the product are signals too
[06:08](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=368s)
Arca's system watches both the external world and activity inside its own product. If Jane provides new information in a meeting, a person can add it to the interface and wake an agent. Rescheduling a calendar event or marking a task complete also creates a signal. Agents communicate with each other as work changes. Singh gives the example of one agent notifying another that the required cash has changed from $100,000 to $80,000, prompting an update to the shared plan.

### The system depends on four operational primitives
[06:52](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=412s)
Singh groups the system around what people need to do at work. Agents need awareness, provided by the signal system. They need to know what to do next, provided by operational context. They need to perform and track work, provided by the harness. They need to preserve state over time, provided by persistent memory. They also need a way to bring in the right people at the right time. The simple outline becomes complicated when hundreds of signals, tasks, agents, and human participants interact.

### Arca improves the agents by changing the firm's operating model
[09:00](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=540s)
During the six months before the talk, Arca worked on closing context gaps, improving the precision of tasks, and deciding whether future work belongs in memory or should become an immediate task. The company also standardised how work happens and encoded that operational context into agent state. Since Arca's own colleagues use the product, their detailed reasoning can become evaluations for improving the system. Singh connects this work to Arca's position as the firm itself, rather than a software vendor selling a separate product.

### Business goals become agent goals that drive employee work
[10:11](https://www.youtube.com/watch?v=oP2WLFpMoDY&t=611s)
Singh says Arca wants agents to drive the business and make the decisions that shape its outcomes. Employees see tasks prepared by agents that have already read the email and reviewed the document, with people pulled in when their action is needed. He describes a future goal of preparing every client's tax situation in January and reaching out before the client asks. If that goal is added to the agents' goals, they can produce analyses and one-click approval emails for employees. In this model, steering the agents is how Arca steers the business.

## Notable quotes
- "What we mean by AI native is that agents run our business." (00:58)
- "We gave them signals from the external world that keeps them really aware of what is happening." (03:13)
- "We let our agents watch the world. So people don't have to." (06:08)
- "At Arca, agents prompt people more than people are prompting agents." (10:45)
- "In an AI native company, you know you've built one when you can steer your business by steering your agents." (12:44)

## Tools & references mentioned
- Arca
- ACH
- Apple

## Who should watch
- You are building agents for work that pauses between emails, approvals, settlements, and other events, rather than a single uninterrupted task.
- Your team needs agents to track operational state and decide when a person should be brought into a process.
- You are deciding whether AI should remain an employee tool or become part of how the business sets and executes its goals.

## Related talks

- [Agents for Everything Else](https://aietalks.com/talks/agents-for-everything-else) (swyx, AI Engineer, 14:10)
- [Building Agents (the hard parts!)](https://aietalks.com/talks/building-agents-the-hard-parts) (Rita Kozlov, Cloudflare, 21:12)
- [Conquering Agent Chaos](https://aietalks.com/talks/conquering-agent-chaos) (Rick Blalock, Agentuity, 14:40)
- [The Agent Works. Does Anyone Trust It?](https://aietalks.com/talks/the-agent-works-does-anyone-trust-it) (Michele Grant, Leading at Scale, 15:05)
- [The Age of the Agent](https://aietalks.com/talks/the-age-of-the-agent) (Flo Crivello, Lindy, 14:49)
