AI-generated text

Arca’s AI-native wealth management: agentic workflows, persistent memory and built‑in auditability

Arca’s head of AI, Aman Singh, outlined how the firm is built from the ground up as an AI-native wealth manager, describing a custom harness, persistent agent memory and workflows that can span days.

AI Newsroom · 31 October 2026 · 4 min read

Arca head of AI Aman Singh spoke with Latent Space editorial head Richard MacManus at AI Engineer New York about how Arca is building an AI‑native wealth management firm. The conversation detailed Arca’s technical choices: a custom harness, persistent memory for agents, and agentic workflows that can run for days rather than just hours.

Architecture: custom harness and task routing

Singh described that Arca has implemented a bespoke architecture rather than merely bolting on AI tools. Central to this is a custom harness that connects models and subsystems, routing tasks to the appropriate component. The routing system ensures a given financial task is sent to the right model or human reviewer.

Long‑running agent workflows and persistent memory

Arca’s agents are designed to handle multi‑day workflows: these are not simple single‑round prompt interactions but ongoing, stateful activities. Persistent memory allows agents to retain prior states and decisions so they can follow and advance complex, multi‑step tasks over time.

Human approvals and concrete examples

A recurring theme was hybrid human–AI collaboration. Singh explained that for certain financial decisions — for example, handling a $100,000 tax bill — agents generate recommendations but human experts must approve final actions. This approach reduces risk while leveraging AI for scale and efficiency.

Safety guardrails and prompt injection defenses

Arca has implemented guardrails and defenses against malicious prompt injection. Built‑in safety mechanisms prevent agents from executing uncontrolled or dangerous commands and ensure recommendations conform to internal policies.

Compliance and auditability by design

Singh emphasized that compliance and auditability are first‑class features: Arca’s systems are designed so every decision and step is traceable, which is critical for regulatory requirements and internal controls in financial services.

Model selection, internal benchmarks and decision models

Arca uses internal benchmarks to select and continuously evaluate models. They also employ decision models to help verify financial facts and make recommendations more consistent.

Why Arca owns its agent infrastructure; AI‑native versus AI‑tool users

Singh explained why Arca owns and operates its agent infrastructure: it provides tighter control, meets specific security and compliance needs, and enables tuning of the platform to business requirements. This differentiates an AI‑native firm from organizations that simply integrate external AI tools.

Sandboxes, verification and human role in trade execution

The firm operates sandbox environments to test changes and verify financial plans. At the same time, Singh noted that trade execution remains with human experts — critical execution steps keep human decision‑making and accountability in place.

Operational feedback, evaluations and client examples

Arca turns staff feedback into evaluations and system improvements, and codifies operational knowledge into agent context. Example client scenarios mentioned include uncovering hidden fees and providing proactive service, illustrating practical applications of the approach.

Takeaway

The conversation framed Arca as an AI‑native wealth manager: a proprietary infrastructure, agents with persistent memory, long‑running workflows, and compliance and auditability baked into the system. The aim is to increase scalability and reduce risk while retaining human expertise at critical execution points.