Agents Prompt People: Using Agents for Long-Horizon Financial Work

What it claims

Arca, an AI-native wealth management firm that raised $48.5 million in a round led by General Catalyst (with Index Ventures and Venrock), runs agents on financial tasks that last up to ten days. Aman Singh, Arca's lead AI Engineer, describes this work as "long-horizon" but not "long-running": it is "very stop-start, there's many actors in between, many decision points, many handoffs." His summary of the design is that "agents prompt people more than people prompt agents."

How it works

Per-client memory tracks five things: what has happened so far, what needs to happen in the next couple of days, what to track over the long term, what to track next, and which individual owns each action. Agents write to this memory.

The safety boundary is drawn by tool access, not by model judgment. Singh says "all of our safety restrictions are fully deterministic." Agents can run tools that carry no client risk. Their role stops at analysis and proposals. Trading and any other money movement go to a human advisor. Arca also runs verification checks for known attacks such as prompt injection.

The agent itself is not sandboxed, but it runs computation in sandboxes. Data access from the sandbox uses a translation step. Code the agent writes is converted into a strict format Arca defines, and the original code is thrown away. Guardrails are applied to that format before anything reaches the backend. This lets agents aggregate and transform client data with code, for example to build a financial plan, without running arbitrary code against the backend.

A meta-agent reads advisor feedback and agent traces and suggests changes to the harness. According to the podcast timestamps, staff feedback also becomes evals.

flowchart LR
  A[Agent task] --> B{Risk to client?}
  B -- no --> C[Agent runs tool]
  B -- yes, money movement --> D[Human advisor decides]
  E[Agent writes code] --> F[Translate to strict format]
  F --> G[Discard code]
  G --> H[Guardrails on format]
  H --> I[Backend data]
  C --> J[Traces]
  D --> K[Advisor feedback]
  J --> L[Meta-agent]
  K --> L
  L -.suggests changes.-> M[Harness]

Arca owns its harness rather than using a managed one. The harness picks a model per task from internal benchmarks, drawing on models from Anthropic (the primary provider, Claude) and OpenAI. Singh gives provider portability as the reason for owning it.

Limits

The page is a short write-up of a podcast, not a technical paper. It gives no benchmark names or scores, no model versions, no description of the strict query format or its guardrails, and no cost or latency figures. The podcast covers topics the text does not expand on: compliance and auditability, decision models for checking financial facts, and a $100,000 tax bill example. All claims come from Arca and are not independently verified.