Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Aman Singh, head of AI at Arca, explains how the firm is building an AI-native wealth management business in which agents run back-office work, monitor clients proactively, and coordinate multi-day workflows, while humans retain advisory relationships and trade execution. The discussion covers model selection through internal benchmarks, a model-agnostic custom harness with tiered routing, persistent memory for long-running tasks, deterministic and LLM-based guardrails, auditability by default, decision models for financial accuracy, and feedback loops that turn staff edits into evals.
Building an AI-Native Wealth Management Firm
Arca is not positioning itself as a SaaS vendor. As Aman Singh puts it, “We’re actually not a SaaS company. We are the firm ourselves.” The company offers a broad range of financial services, including financial advice and planning, tax planning, insurance, and estate planning, for clients ranging from newly wealthy founders who have just sold a business to large technology executives planning early retirement and young families buying their first home. The core design principle is that AI replaces the work surrounding the advisor, not the advisor. Clients never interact with AI directly; they experience a level of service delivered through the empathy and care of a human advisor, while agents run back-office work and monitor each client’s situation continuously, including market movements and the life events that affect their finances.
The phrase “Agents prompt people more than people prompt agents” captures the shift. In a tools-based services firm, users prompt AI to produce work. At Arca, agents initiate proactive, 24/7 service and pull in the right people at the right moments.
Models, Harness, and Routing
Arca uses frontier models from Anthropic and OpenAI, with Claude as the primary model. The company has moved through several providers over time, previously using Gemini before OpenAI and then Anthropic. Model selection is driven entirely by internal benchmarks that test the specific capabilities the harness must express when running the business. External benchmarks are not the deciding factor: a recent model that looked strong externally proved too slow and too expensive in practice. Arca’s principle is to remain model agnostic and avoid dependence on any single provider or lab’s harness, keeping optionality open.
The harness is built in-house and is model agnostic. Routing works through tiers of models matched to task complexity. A tax plan or tax analysis draws on the heaviest tier, while the large volume of background activity runs on lighter tiers. Advisors are unaware of these routing decisions, which are abstracted away entirely.
Long-Running Workflows and Memory
Many of Arca’s workflows span roughly ten days, which is far longer than most agent deployments. Consider a client tax bill that requires liquidating assets to raise $100,000. An agent handles the analysis of where the client’s money is allocated and the most optimal way to liquidate. Execution is different: the agent can access the custodian’s APIs but cannot move money, liquidate assets, or communicate directly with clients without human-in-the-loop approval. Agents assemble a plan for the relevant human expert to review and adjust, then pull in the right person to execute.
Continuity across days depends on memory systems that every agent contributes to. For each client, the memory tracks what is happening now, what is being serviced, who owns which action, what has already happened, what is tracked for the near and long term, and which individual is responsible for each step. A daily reflection process, described as a kind of dreaming, reviews the state of the world for each client and updates memory to persist long-horizon tasks. When a new signal arrives, such as an unattached phone call, a fresh agent wakes up, ramps on the client’s situation from memory, and determines whether the signal relates to an ongoing process like the tax bill. Signals that belong to an existing thread, such as a reply in an email chain, route to the agent already reading that thread.
Routing people into a process follows the same logic. A step seven actor may not need to know about a process until the right moment. Agents use long-term memory and the daily reflection to determine when a signal implies a particular person must be involved. Arca has encoded how the industry works and how Arca wants work to happen, so agents can participate in standardized procedures.
Safety, Guardrails, and Verification
Safety restrictions are deterministic and take the decision away from the agent. Some tools execute directly, such as simple record keeping. Every tool call that writes data passes through fast LLM-based verification. External inputs such as emails, documents, and phone calls are treated as prompt injection surfaces, and verification runs at those edges to detect injection patterns. Any data leaving the system, including web searches and calendar invites, faces deterministic PII checks and a consensus of small models screening for exfiltration. The guiding principle is to contain agents at the edges of capability rather than earlier.
Financial plans follow a specific verification path. Agents query data through code in a sandbox, which is one harness component rather than the agent’s home environment. Code output is translated into a strict format, the code is discarded, and guardrails apply before anything reaches the backend. The sandbox produces client-facing artifacts such as PDF financial plans. Before publication, deterministic checks catch things like raw account numbers, and smaller models flag non-compliant language, such as promising returns. The team argues that AI improves compliance here, since it is tireless at catching even minor non-compliant phrasing.
Compliance and Auditability
Industry regulation cares about outcomes, but compliance cares about auditability, and Arca builds auditability in by default. Every action produces a change log recording what changed, who changed it, from which part of the system, and when. This applies to human actors too: a UI change affecting a client is logged. Beyond regulatory compliance, the same change tracking feeds the agentic loop, since a UI change can inform an agent that something happened.
Owning the Intelligence
Arca is bullish on frontier models but cautious about depending on managed agent products. Those products often own how models persist memory and operational knowledge. For Arca, that knowledge is the business’s intelligence, and handing it to another provider means losing the control needed to build something unique and tailored enough to affect the business. Anthropic moving into financial services is seen as validating for the industry, but the distinction is clear: a services firm with access to AI tools differs from a firm built to be AI native.
Decision Models and Financial Accuracy
Models working with financial data tend to be confidently wrong. A typical example is applying 2025 tax brackets to a current analysis. Arca uses decision models within the harness to detect when a model is about to make a financial recommendation or take a financially consequential action. The model is then nudged toward a knowledge base of verified current facts or to web search. Owning the harness makes this kind of intervention possible.
Trade execution remains human. Autonomous trading is not in place. Agents propose trades, and experts decide whether to make and execute them. Recovery from a failed trade is not handled by decision models today.
Feedback Loops and Evals
Staff edits and cancellations become evals. Feedback is stored alongside the traces of the agents that produced it. Every night, a meta agent analyzes traces from real-world operation and suggests harness changes, such as misleading tools, errors, or gaps in the knowledge base. A human engineer still reviews prompt changes before they ship. The team has observed that agents making prompt changes tend to overfit each fix into a sprawling prompt rather than generalizing it, which is why human review remains in place.
Operational Knowledge as Context
Arca maintains a knowledge base of files that serves as operational context for agents. Experts curate and update it, and it is a first-class feature of the product, visible as a knowledge tab. Agents cite knowledge as they work. When they find it incorrect, the issue surfaces through trace analysis, and people can edit the underlying skill directly.
Client Outcomes
In one recent case, a prospective client already had an adviser charging $1,500 for services that were not responsive. Agents read the client’s documents, analyzed stated goals and the current adviser’s statements, and found fees the client had not been aware of. They also established that the adviser was not a fiduciary and therefore had no obligation to act in the client’s interests. Agents also proactively flag deadlines and risks, such as an approaching required minimum distribution, a task that is difficult for humans to track at scale but is the default behavior of Arca’s agents.