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

Arca is an AI-native wealth management firm, not a SaaS company selling tools to other firms. Aman Singh, who leads the AI platform team, explains that the firm’s agents do not replace financial advisors; they replace all the work around the advisor. Clients never interface with AI directly. Instead, they experience a level of proactive, responsive service that the industry has not previously delivered, mediated through human advisors. The core design principle is that agents prompt people more than people prompt agents.

The Harness, Not the Model, Is the Product

Arca uses frontier models from Anthropic (primarily Claude) and OpenAI, but the firm is deliberately model-agnostic. Internal benchmarks, not external leaderboards, guide model selection. For example, despite strong public benchmarks, Llama was rejected because it was too slow and too expensive for Arca’s use case. The critical piece of infrastructure is the internally built “harness” that abstracts model choice away from both advisors and end users. The harness assigns tasks to different tiers of models based on complexity: a heavy tax analysis gets the most capable model, while routine background monitoring uses a cheaper, faster one. The key insight is that owning the harness means owning the intelligence that differentiates the business. Singh explicitly avoids dependencies on products like Anthropic’s mana