Using Agents for Long-Horizon Financial Work — Aman Singh, Arca

Aman Singh, who leads Arca’s AI platform team, says Arca is building the infrastructure needed to “run the business with AI.” The firm provides a broad range of wealth-management services: financial advice and planning, tax planning, insurance, and estate planning. Its clients include people who have sold a business and must coordinate complicated financial decisions, technology executives planning to retire early, and families trying to buy their first home. Singh emphasizes that Arca is not using AI to replace financial advisors. Instead, AI replaces the work around the advisor: analyzing markets, reviewing documents, tracking a client’s changing circumstances, and preparing recommendations. Clients generally do not interact directly with the AI. They experience what Arca describes as a more proactive and responsive service delivered through a human advisor’s empathy and care. The system is constantly considering questions such as what is happening in the market, how that affects a particular client, and what “textures” of that client’s life may change their financial needs. Singh contrasts Arca with an ordinary financial-services firm that merely adds AI tools. Arca is the firm itself, not just a SaaS provider selling software, so it can standardize how work is done and make agents proactive. His central formulation is that “agents prompt people more than people prompt agents”: the system runs continuously and brings humans into a process when their judgment or action is needed.

Arca currently uses frontier models from companies including Anthropic and OpenAI. Claude has been its primary model, although the team previously used Gemini and OpenAI models. Singh says external benchmarks do not determine these choices. Arca runs internal benchmarks covering the capabilities its general-purpose harness needs in order to operate the business. He gives Luna as an example: despite favorable external benchmarks, it was too slow and expensive for Arca’s workloads, so the internal tests led the team away from it. Arca built its own model-agnostic harness rather than adopting a separate harness tied to one model provider. Singh says the goal is to preserve optionality and avoid depending on a particular lab or orchestration product. The harness assigns models in tiers according to task complexity. A tax plan or tax analysis receives the heaviest tier, where Arca might use an Opus-level model, a Sonnet-level model, or a heavy GPT model. Less demanding background work is handled by lower-tier models. Financial advisors do not choose models while using the product, and they are not expected to know which model is handling a task. That selection is abstracted away inside Arca’s system. Singh later makes the same argument about Anthropic’s managed-agents infrastructure: Arca is bullish on Anthropic’s models, but does not want an outside provider to own the persistence, operational knowledge, and intelligence that determine how its business runs.

Singh describes a portfolio-adjustment workflow that can run for roughly ten days. It begins when a client emails a tax bill. Agents analyze where the client’s money is currently allocated and determine the most suitable way to liquidate enough assets to produce, for example, $100,000. That analysis is fully automated and is considered an appropriate use of agents. The agents cannot independently move money, liquidate assets, or execute the trade. Arca has a deep API integration with its custodian, but it places strict human-in-the-loop guardrails around financial actions and direct communication with clients. Once the analysis is complete, the agents bring in the appropriate person, provide a proposed plan, and allow that person to review and adjust it before execution. Singh clarifies that Arca is not currently doing autonomous trading: agents propose trades, while human experts decide whether the trades are appropriate and execute them. Long-running work is maintained through client-specific memory systems. These record what is happening, what the firm is trying to service, who owns each action, what has already happened, and what is due immediately, soon, or in the longer term. A continuing email thread can reactivate the same agent, while a new signal—such as a phone call—can wake a fresh agent. That agent uses the stored state and a recurring reflective process to understand how the new signal relates to the client’s ongoing tax or investment situation.

Arca’s safety restrictions are designed to remove key decisions from the agent rather than trusting the agent to follow a verbal rule. Simple record-keeping tools may be executed directly, but every tool call that writes data is checked by an additional language-model verification step. The verifier analyzes each database write. Arca also checks incoming emails, documents, and phone-call information for prompt injection. The system applies further checks whenever data leaves Arca. For web searches, calendar invitations, or other external actions, deterministic rules look for personally identifiable information, while a consensus of smaller models checks for attempted data exfiltration. Singh says the principle is to place guardrails at the edges of the agents’ capabilities rather than constraining every internal reasoning step. Compliance depends heavily on auditability. Every agent action produces a change log recording what changed, who or what changed it, which part of the system made the change, and when. The same logging applies to humans: a staff member changing a client-affecting field in the user interface generates an equivalent record. This gives Arca an audit trail for regulators and also supplies agents with signals about changes made by people. Singh says compliance can sometimes be stronger than a human-only process because AI can repeatedly check an artifact for small instances of noncompliant language, such as promising investment returns or exposing a raw account number.

Arca uses sandboxes as a component of its harness, but the agent itself does not live inside the sandbox. Instead, the agent sends computation there, where code can query and manipulate financial data to produce an output such as a client’s financial plan. Arca built a translation layer that converts the code’s result into a strict format, discards the original code, and applies guardrails before the result reaches the backend. The sandbox can generate a client-facing artifact, including a PDF explaining the proposed financial plan and how Arca intends to approach the client’s finances. That artifact is not immediately publishable. Arca runs deterministic checks for prohibited content, including raw account numbers, and uses smaller models to identify potential compliance problems. One example is language that promises a particular return. Singh also explains how Arca uses smaller “decision models,” including a model he calls Jev, to improve financial accuracy. General models can be confidently wrong when they handle time-sensitive financial facts. For instance, a model may use 2025 tax brackets for a current analysis without realizing that the figures are outdated. Arca’s decision layer detects when a recommendation requires financial rigor and nudges the main model to consult Arca’s verified knowledge base or perform a web search. Decision models are not currently used to recover from failed trades, because agents do not execute trades autonomously in the first place.

Arca treats staff edits and cancellations as feedback rather than simply as manual corrections. Each piece of feedback is stored alongside the trace of the agent that produced the original work. Every night, Arca collects traces from agents operating in the real world and runs a meta-agent over them. That process suggests improvements, such as fixing a misleading tool, addressing an error, or adding missing information to the knowledge base. A human engineer still approves prompt changes before they ship. Singh says agents make “really nasty prompt changes” when left unsupervised: instead of applying a narrow, general fix, they often replace the entire prompt. The current process therefore uses agents to identify likely improvements while retaining a person as the final gatekeeper. Arca’s operational knowledge is represented as a collection of files in a first-class product area called Knowledge. Experts curate and update these files directly. Agents can cite the knowledge they use, and trace analysis can reveal when the underlying guidance is wrong or incomplete. This creates a feedback loop: the knowledge is visible to staff, staff can edit the relevant skill, and the resulting changes can improve future agent behavior. Singh’s example of the practical benefit involves a prospective client who already paid another advisor $1,500 but complained that the advisor was unresponsive. Arca’s agents reviewed the person’s statements, investments, goals, and existing advisor relationship. They found fees the person did not know about and identified that the advisor was not a fiduciary. Singh says this kind of persistent attention can also catch deadlines, such as a client being at risk of missing a required minimum distribution, at a scale that is difficult for humans to manage manually.