Arca’s AI‑Driven Wealth Management Uses Persistent‑Memory Agents for Multi‑Day Client Work

Long‑Horizon Agent Workflows

At the AI Engineer New York conference, Aman Singh, lead AI engineer at Arca, explained that the firm’s agents can handle tasks lasting up to ten days.

He described the work as “very stop‑start, there’s many actors in between, many decision points, many handoffs.”

To keep track of such extended processes, Arca built an internal memory system that records every client interaction, upcoming actions, and long‑term goals.

The memory logs also assign responsibility for each step to a specific individual, ensuring accountability across the workflow.

Agents contribute continuously to this persistent record, allowing the team to query “what has happened so far” and “what needs to happen next.”

Model Selection and Custom Harness

Arca relies on frontier‑lab models from Anthropic and OpenAI, with Claude serving as the primary provider.

Rather than using a generic managed harness, the company built its own harness that automatically selects the optimal model for each task based on internal benchmarks.

This ownership gives Arca the flexibility to switch model providers without disrupting the agent pipeline.

The harness is further refined by a “meta‑agent” that analyzes traces from both agents and human advisors, then suggests adjustments to improve performance.

Safety Controls and Human Oversight

All safety restrictions in Arca’s system are fully deterministic, meaning agents never make final trading or money‑movement decisions.

Agents are limited to analysis and proposing solutions; a human advisor executes any transaction that carries client risk.

Arca runs verification checks for known AI security issues such as prompt injections.

While the agents themselves are not sandboxed, the code they generate runs inside sandboxes that translate the output into a strict, pre‑approved format before reaching backend systems.

This extra layer allows agents to write code that queries, aggregates, and manipulates client data to produce financial plans, while still protecting sensitive information.

Human advisors also provide feedback on agent suggestions, which is fed back into the meta‑agent to continuously improve the harness.

Arca recently secured $48.5 million in funding, led by General Catalyst with participation from Index Ventures and Venrock, underscoring investor confidence in this AI‑native approach.

Singh summed up the relationship between agents and staff as “agents prompt people more than people prompt agents.”

By combining persistent memory, a custom model‑selection harness, and deterministic safety guards, Arca aims to scale complex, multi‑day advisory work while keeping ultimate decision authority with human professionals.

Why This Matters: the persistent‑memory agent framework lets Arca scale long‑horizon advisory work while keeping final decisions and risk‑bearing actions in human hands.