Source: latent.space/p/using_agents_for_long-horizon_financial_work_aman_singh_arca
Arca, an AI-native wealth management firm, uses agents for most background work, while human advisors make the final calls. Aman Singh, the company's lead AI engineer, said agent tasks at the firm can take up to ten days, with many actors, decision points and handoffs. Arca recently raised $48.5 million in a round led by General Catalyst.
To keep context across that span, Arca built an internal memory system that records each client's history, upcoming steps and which person owns each action. The company runs its own harness, which picks the model for each task based on internal benchmarks. Claude is its primary model provider, and owning the harness lets Arca switch providers if it needs to.
Safety rules are fully deterministic and keep agents to analysis and proposals, with trading and money movement left to humans. Agents can run code in sandboxes to query client data, but that code is translated into a strict expected format and checked by guardrails before it reaches back-end systems. A meta-agent reviews advisor feedback and agent traces to suggest changes to the harness.
Why it matters if you build agents
Teams building long-running agents may need persistent memory that tracks owners and pending actions, not just conversation history. Arca's setup shows that hard limits enforced in code, rather than in prompts, can keep agents away from high-risk actions. Owning the harness also keeps model choice open as providers change.
Get stories like this every weekday
One short email with the day's agent tooling news.