Long-horizon agents run a business by turning a client goal into days of stop-start work rather than minutes of token grinding: a signal system, operational context, a tool harness and persistent memory let the agent own an outcome, and people step in only when a decision needs a human. Arca, an AI-native financial services firm, applies this to tax planning across roughly ten days of hand-offs.
What long-horizon agents mean in an AI-native firm
Long-horizon agents are software workers that own a multi-day business outcome end to end, waking on new events rather than answering a single prompt. Aman Singh, who leads AI at Arca, an AI-native financial services firm, drew that line between an agent that assists work and an agent that executes it.
Arca's definition of AI native is narrow and testable: agents run the business, and they do it outside a code editor. Singh contrasted that with tools like GitHub Copilot, where an assistant sits beside a person doing the work. At Arca the agents carry the client outcome.
The shape of financial services work explains the design. Singh described it as long horizon and less long running. One client goal may span ten days with advisors, traders and operations staff, settlement delays and incoming messages between each step. A firm doing that repeatedly spends more coordination time as client count grows, which is the pressure that sent Arca toward agent ownership.
Arca is not a SaaS vendor selling software to other firms. Singh described an applied AI company that happens to run a services business, which matters because it lets the firm change its own operating procedures and then encode them as agent context. Watch the full talk on the AI Engineer channel for Singh's own walkthrough.
The five primitives Arca gives its agents
Arca gives its agents five capabilities, and Singh framed them as the things any colleague needs: awareness of events, knowledge of what to do next, tools to act, memory to hold state, and access to other people. Each maps to a separate piece of the platform rather than to a single prompt.
The table summarizes the split and what each primitive is for. Sources are Singh's own description of Arca's system; no external benchmark of the stack has been published.
| Primitive | Arca component | What it lets an agent do |
|---|---|---|
| Awareness | Signal system | Read emails, calls, custodian events and market moves |
| Knowledge | Operational context | Know how the firm handles a tax or settlement case |
| Action | Tool harness with sandbox | Run financial computation and execute steps |
| State | Persistent memory | Hold client facts and work tracked for later |
| Collaboration | Shared spaces | Pull the right person in at the right moment |
Singh noted that harnesses across the industry are coalescing on a similar shape, so he spends little of the talk on the harness itself. The interesting design sits outside it, in how work is routed between agents and people. The AI Engineer conference hosted the talk in October 2026.
Why one tax email can take ten days to resolve
A single tax email can take ten days because each step in the chain depends on settlement windows and on other people acting. Singh's example is a client named Jane with a large tax bill due April 15, and the goal is raising cash in the most tax-efficient way.
The sequence is inherently stop-start. An advisor analyzes statements and proposes the tax-optimal sale. A trader liquidates the assets. Settlement takes days. Operations staff then trigger an ACH transfer, which also takes days to land. Every stage waits on the previous one, and every stage involves a different person.
Singh's point about scale is the practical one. Run that workflow for several clients at once and the firm's coordination time grows faster than the client list. The bottleneck sits in the hand-offs between specialists. One useful reference for how agent tooling is being framed across the industry is the Model Context Protocol, an open standard for connecting models to tools and data.
Arca's answer was to remove people from the critical path without removing them from the decisions. Agents own the goal for its full duration and surface work only when a human signature or judgment is genuinely required. Singh said nobody has to babysit the goal across those days.
How the signal system keeps agents aware without babysitting
A signal system keeps agents aware by turning every event that matters into a message that interested agents receive. Singh listed the inputs: emails, phone calls the firm makes, documents arriving, meetings, custodian events, and market movement. Background agents evaluate all of it continuously.
The mechanism is subscription rather than polling by a person. Agents register interest in specific future events, such as a client replying or a transfer landing, and the signal layer routes the event to the agents that asked for it. Internal product actions count too. Rescheduling a calendar event or marking a task complete is itself a signal.
If an expected event never arrives, agents can wake themselves on a schedule to check on stalled work. In Singh's story the client never answers the email, and a colleague later learns something in person that changes the plan. That update goes in through the product interface, which wakes the relevant agent.
A signal can also move work between agents. When the plan changes, a background agent can notify a second agent to update the shared space, for example revising a required amount from $100,000 down to $80,000.
Operational context and memory: what the agents know
Operational context and memory give agents the knowledge to act and the state to remember it later. Singh treats them as separate systems: context covers how the firm handles a situation, while memory covers durable client facts and work deferred to the future.
The memory example he gave is concrete. An agent should never sell a client's Apple position, because the client said so on a call eight months earlier. That kind of constraint has to survive across sessions, staff changes and the many days a goal runs.
Persistence is the part Singh said the firm has spent months tightening. Work that can be done now becomes a task for a person or an agent. Work that must wait is tracked in memory until the triggering event arrives. Arca has been closing context gaps so agents know enough to reach the correct next step without asking.
Because Arca is the firm, it can also standardize its own procedures first and then encode them. That is a different position from a software vendor trying to infer a customer's process from the outside.
Shared spaces where agents pull people in
Shared spaces are the surface where agents and people work on the same plan, with the agent asking for a specific action at a specific moment. Singh described the agent building a space, posting its plan there, and bringing in the right colleague when execution needs a human.
Multiple people can chat with the agent inside the same space. The agent approves the plan and the outbound email, then tracks whether the client has replied and what the reply means for the plan. Later steps route to whoever holds the relevant role, such as the trader who handles those assets.
Singh's measure of success inverts the usual interface pattern. At Arca, agents prompt people more than people prompt agents. When a person logs in, they see a task list prepared by agents that have already read the email, reviewed the document and done most of the work.
The design also produces training signal. Colleagues are the product's users, so their reasoning as they work becomes evaluation material for improving the system. Singh cited that feedback loop as a practical advantage of operating the business rather than selling to it.
How Arca steers the business by steering its agents
Arca steers the business by adding a goal to its agents' goal set, then letting them drive work through people toward it. Singh's example is a proactivity goal: prepare every client's tax situation each January and reach out before anyone asks.
The tax email from the opening acts as a failure case. Jane should never have needed to send it, because a proactive firm would have started the analysis first. Once proactivity becomes an agent goal, the same primitives handle it: analyses run in January and approval emails arrive on users' plates with a full plan attached.
Singh's stated line between a company that uses AI tools and an AI-native one is whether the business can be steered through its agents. That is a claim about operating model, not a published performance result. No independent measurement of Arca's throughput, accuracy or cost has been released.
The talk sits alongside a wider discussion of agent harness design on the AI Engineer YouTube channel. Whether the pattern transfers to firms without Arca's ability to standardize its own process is the open question.
Frequently asked questions about long-horizon agents
- What are long-horizon agents? Long-horizon agents are software workers that own a business outcome across days of stop-start activity, holding state and waking on new events. Arca uses them for client goals like raising cash for a tax bill, where specialist hand-offs and settlement delays separate each step.
- What is the difference between long-horizon and long-running agents? Long-running agents grind on tokens for hours or days without interruption. Long-horizon agents wait, resume and route work across many actors and decisions, and the total elapsed time comes from settlement windows and human availability rather than compute time.
- Why does Arca use a signal system instead of polling? A signal system lets agents register interest in specific future events and receive them when they occur. It covers external inputs like emails and custodian events plus internal actions like task completion, so agents stay aware without a person checking a queue.
- Do people still make decisions in Arca's workflow? People approve plans, execute trades and supply information the agents cannot obtain. The agents do most of the preparation and pull a person in when a signature or judgment is needed, which is why agents prompt people more often than the reverse.
- Is Arca a software vendor? No. Arca runs a financial services business and builds its own agent platform internally. That lets the firm change its own procedures and encode the changes as agent context instead of inferring a customer's workflow.
From transcript to published article
Singh's talk is a worked example of a harder pattern: knowledge that lives in a conference video, a podcast episode or a recorded interview and never becomes findable text. If you have recorded explanations, client stories or technical walkthroughs sitting on YouTube, that material can become a written article instead.
If you are learning to build software in this style, Crazystack Typescript and the Bootcamp do Dev Doido are worth a look for structured, project-based practice, and Gustavo Dev Doido covers related material in Portuguese. The same instinct applies to writing: turn what you already recorded into something a reader can search and quote.
Turn your recorded knowledge into a searchable article
The primitives that let an agent own a ten-day goal are the same ones that make a recorded conversation useful: awareness of what was said, context about why it matters, and memory of the details. A video already holds all three. It just needs a written form.
Paste a YouTube URL into Skala Blog, let the tool transcribe the video, and generate a structured draft you can review and edit before publishing. The flow is the same one behind this piece, published in 2026: recording, transcription, article.
The pipeline, step by step:
- Record the conversation.
- Transcribe it into text.
- Generate a structured draft.
- Review and edit before publishing.
| Stage | Input | Output |
|---|---|---|
| Recording | Conversation | Video |
| Transcription | Video | Text |
| Drafting | Text | Structured draft |
| Review | Draft | Published article |
The tool has run this same flow since 2026, when this piece went through it.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits