Aman Singh on AI Agents for Long-Running Financial Work at Arca

At AI Engineer New York, Latent Space editor Richard MacManus interviews Aman Singh, who leads the AI platform team at Arca. Singh says the wealth-management firm is building an internal agent harness to support the business, not selling a standalone software product.

Arca offers financial advice and planning across investments, tax, insurance and estates. Singh describes clients ranging from a business owner who has sold a company, to a technology executive planning early retirement, to a family preparing to buy its first home.

Singh says AI is intended to replace work around the advisor, not the human advisor. Clients interact with an advisor, while agents carry out background work intended to make service more proactive and responsive. He presents this as Arca's operating model, rather than a guarantee of financial outcomes.

Arca currently uses frontier models from providers including Anthropic and OpenAI; Claude has been its primary model. Singh says the company has changed providers over time and uses its own internal benchmarks, covering the capabilities its harness needs, to evaluate candidates.

External benchmark results do not always predict Arca's experience. Singh describes a model that performed worse than expected on the firm's internal tests in speed and cost. He wants the architecture to remain model-agnostic so the company can retain a choice of providers.

The firm has built its own model-agnostic harness. Tasks are routed to tiers according to complexity: a tax analysis may use a heavier model, while less demanding background activity uses a lower tier. Advisors interact with the product without making those model-selection decisions themselves.

MacManus asks about an example involving a client's $100,000 tax bill. Singh says a workflow may span ten days. Agents can analyze where the client's assets are held and propose how to liquidate enough to meet the bill, while other steps require people.

Arca's agents do not autonomously move money, liquidate assets or communicate directly with end clients, Singh says. Although the firm has custodian APIs that could support transactions, a human reviews and can adjust the proposed plan and executes the trade when appropriate.

For work that resumes days later, the system receives different kinds of signals. A reply in an existing email thread can return to the agent following it; a new phone call may wake another agent that must determine how the call relates to ongoing work.

Arca maintains persistent client context: what is underway, what has happened, who owns each action and what needs attention soon or later. Agents contribute to this memory, and a periodic reflection process updates it so a newly activated agent can understand the relevant state.

Singh says critical safety restrictions are deterministic: the agent does not decide for itself whether a protected action is permitted. Routine record-keeping tools may be available directly. Separately, the company uses fast LLM-based checks on database writes as an additional verification layer.

Incoming emails, documents and calls can carry prompt injection. Singh says Arca checks inputs at the edges, and checks information leaving the system, such as through web searches or calendar invitations. Those checks include deterministic detection of known personal information and small-model consensus intended to detect exfiltration.

For compliance, Singh emphasizes auditability. He says every agent action produces a change log recording what changed, who or what changed it, where and when. The same logging applies to human edits in the interface and can also inform agents that the state has changed.

The interview returns to Singh's phrase that agents prompt people more often than people prompt agents. In a multi-day process, not every participant needs to be involved immediately. Persistent context and daily reflection help an agent identify the appropriate person when a relevant signal arrives.

Singh says Arca encodes not only general industry practice but also how it wants its own work to happen. Because it operates the advisory firm itself, it can standardize workflows and let agents participate in those processes rather than adapting a generic tool to a separate business.

Asked about managed agent infrastructure, Singh is enthusiastic about frontier models but cautious about depending on a managed-agent product. In his view, control of memory and operational knowledge is closely tied to control of a firm's intelligence and its ability to tailor the system to its business.

Singh sees a distinction between an existing services firm that equips staff with AI tools and a firm designed around AI-enabled workflows. He characterizes Arca's agents as proactive background participants that surface work to people, rather than waiting for every task to be prompted by staff.

A sandbox is one component of Arca's harness; the agent itself does not run entirely inside it. Singh says agents offload computation to the sandbox. Code that requests access to data is translated into a strict expected format, with the original code discarded and guardrails applied before the request reaches the backend.

Agents can use code to aggregate data and prepare artifacts such as a financial-plan PDF. Once an artifact is produced, the system verifies it: deterministic checks can catch raw account numbers, while smaller models look for other compliance issues, such as a promise of investment returns.

Singh believes tireless automated checking can improve compliance review, while describing it as part of a layered system. These are the firm's claims about its own process; he does not present a measured comparison demonstrating that automated checks always outperform human review.

Financial models can confidently use stale facts. Singh gives the example of applying an outdated year's tax brackets. Arca uses decision models to detect moments requiring financial rigor and nudge the agent toward a verified, current knowledge base or web search.

The firm is not currently using a decision model to recover from failed trades, Singh says, because its agents do not perform autonomous trading. They may propose trades, but human experts decide whether a trade is appropriate and carry it out.

Staff edits and cancellations provide feedback. Singh says the product collects that feedback together with the trace of the agent run. Each night, a meta-agent examines real traces and suggests improvements, such as fixing misleading tools, errors or gaps in the knowledge base.

A human engineer still reviews prompt changes before they ship. Singh has found that agents left alone may keep adding narrow fixes until a prompt becomes bloated, instead of generalizing the underlying lesson. The feedback loop is designed to produce evaluations and improvements, with human review at release.

Operational knowledge is stored as files and exposed as a product feature that domain experts can curate. Agents can cite the knowledge they use; trace analysis can surface incorrect information, after which staff can edit the relevant skill or knowledge directly.

Asked for an example of client impact, Singh describes a prospective client who already had an advisor. Arca's agents reviewed statements, investments, goals and the stated advisory charge, and surfaced additional fees that the person had not understood, as well as a question about the advisor's fiduciary status.

Singh also gives upcoming deadlines and required minimum distributions as examples of events agents can flag. He says tracking such details at scale is difficult for people and presents proactive monitoring as a default behavior of Arca's system.