Arca's Aman Singh maps AI agents that run for days

Arca's setup, as described by its head of AI Aman Singh at AI Engineer New York, centers on a custom harness and persistent memory rather than relying solely on off-the-shelf agent frameworks. That combination lets the firm run agentic workflows that span days, not hours, according to Singh. He also described choosing models with internal benchmarks and building routing for tasks, suggesting Arca treats model selection as an internal capability rather than an external given.

The human role stays explicit. Trade execution remains with human experts, and financial plan verification is checked in sandboxes. Singh cited handling a $100,000 tax bill as an example involving agents and human approvals, and he noted that agents prompt people more than people prompt agents, which points to a workflow where AI initiates steps and people respond. Compliance and auditability are built in by default, and safety guardrails address prompt injection defenses.

Arca also owns its agent infrastructure and turns staff feedback into evals and improvements, while operational knowledge becomes agent context. Singh drew a contrast between AI-native firms and firms using AI tools, framing Arca's approach as one where agents and human advisors work in a persistent loop rather than as a one-off assistant.