At AI Engineer New York this month, Aman Singh, lead AI Engineer at Arca, outlined how the AI-native wealth management firm combines agent-based automation with human oversight. Arca’s architecture delegates much of the background work to agents, but human advisors retain final authority for decisions that move money or carry client risk.
What agents do and what humans keep control of
According to Singh, agents conduct analysis, develop proposals and prepare next steps, but they do not execute trades or authorize funds transfers. Singh summarized the relationship by saying agents "prompt people more than people prompt agents." Many of the firm’s tasks are long-horizon — they can span multiple days and, Singh noted in his AIE presentation, some agent tasks at Arca can take up to ten days. These processes are often stop-start with multiple actors, decision points and handoffs, which led Arca to build systems to persist state across time.
Internal memory and task tracking
Arca maintains an internal memory for every client that records what has happened so far, what is being tracked for the near term (the next few days), what is being tracked for the long term, and which individual is responsible for which action. Agents contribute to this memory so the platform can follow prior steps and outstanding tasks through multi-day workflows.
Models, a custom harness, and model selection
Arca uses models from frontier labs including Anthropic and OpenAI, naming Claude as its primary model provider. The company has built its own harness — a control layer that automatically selects the best model for a given task based on internal benchmarks. Owning the harness rather than using a managed agent harness gives Arca flexibility to change model providers if needed.
Safety: deterministic guardrails and human execution of money movement
Safety is central for financial services. Singh emphasized that while agents perform much of the analytic work and propose solutions, their role is limited: decision-making authority for trades and any money movement remains with human advisors. Singh described Arca’s safety restrictions as "fully deterministic," meaning predefined rules remove certain decisions from the agents’ purview.
Arca also runs verification checks against known AI security risks such as prompt injections. The agent itself is not fully sandboxed, but sandboxed environments are used for computation.
Secure code execution and data handling
To protect sensitive data, Arca has implemented further security layers for code that runs inside sandboxes. Singh explained they translate code written by an agent into a strict, expected format, discard the original code, and then install guardrails on that format before it reaches backend systems. This design lets agents use code to query, aggregate and manipulate data to produce outputs like financial plans while preventing uncontrolled access to raw data.
Using feedback and a meta-agent to improve the harness
Arca collects feedback from human advisors and traces of agent activity, and uses a "meta-agent" to analyze this data and suggest improvements to the harness. This process turns operational knowledge and staff feedback into evaluations and changes that refine routing, model selection and other behaviors.
Funding and context
Arca recently raised $48.5 million in a financing round led by General Catalyst, with participation from Index Ventures and Venrock. The company is using that capital to continue developing its AI-based wealth management products and infrastructure.
Additional material
The AIE presentation and podcast include a full breakdown of topics and timestamps that cover model choice, harness design, a $100,000 tax-bill example showing agents plus human approvals, memory across long-running tasks, sandboxing and verification, compliance and auditability, and client examples such as uncovering hidden fees and delivering proactive service.
