How a $64M Firm Runs Wealth Management on AI Agents

The setup. Arca is an AI-native wealth management firm that raised $64 million. Aman Singh leads their AI platform team, building a harness that runs the business with AI. The firm does not replace financial advisors but automates the work around them: analysis, research, monitoring, and compliance checks. Clients interact only with human advisors who are supported by a proactive agent system running 24/7.

The architecture. Arca uses frontier models from Anthropic and OpenAI, primarily Claude, but stays model-agnostic through internal benchmarks. They recently tested Luna and found it too slow and costly despite strong external benchmarks. Model selection is tiered by task complexity: tax analysis uses the heaviest tier, background monitoring uses lighter models. The harness itself is custom-built and manages memory, tool execution, and guardrails.

The long-horizon challenge. A client tax-bill scenario spans 10 days. Agents handle internal analysis fully autonomously but cannot execute trades or move money without human approval. Memory systems track every client's state: what happened, who owns what, what is due next. A 'dreaming' process reflects on the world state and updates memory for long-term tasks. New signals, like a phone call, wake a fresh agent that reads the persisted context and contributes to the ongoing process.

Safety and compliance. All safety restrictions are deterministic, removing decision-making from agents. LLM-based verification runs on every write to the database. Prompt injection detection runs on incoming emails, documents, and phone calls. Data leaving the system, such as calendar invites, is checked for PII exfiltration by both deterministic rules and a consensus of small models. Every action produces an audit log, which also powers agentic capabilities.

Feedback and improvement. Every staff edit or cancellation becomes an evaluation. Feedback is coupled with agent traces, and nightly a meta-agent analyzes them to suggest harness improvements: misleading tools, errors, or missing knowledge. Prompt changes still require a human engineer to review because agents tend to over-correct. Knowledge bases are curated by experts as a first-class product feature, with citations and editing built into the advisor workflow.

The bottom line. Arca's agents are more proactive than reactive: they prompt humans, not the other way around. They discovered hidden fees and a non-fiduciary status for a prospective client, something a human might miss. Singh argues that being the firm itself, not a SaaS vendor, lets them standardize workflows and own the intelligence. He warns against depending on managed agent products like Anthropic's, because losing control of memory and operational knowledge means losing the ability to differentiate.