Arca's agents work on wealth tasks that run up to ten days, but every money move stays with a human advisor

"Agents prompt people more than people prompt agents." That line from Arca's lead AI engineer is the best summary of a good business agent I have heard this year.

What happened

Latent Space interviewed Aman Singh, lead AI engineer at Arca, an AI native wealth management firm. FinTech Global reported in June that Arca oversees more than $1bn in client assets with 28 people, after a $48.5m Series A led by General Catalyst.

Singh said agent tasks at Arca can take up to ten days. The work is stop start, with many people, decision points and handoffs. So Arca built a memory system with persistence for every client: what has happened, what comes next, what is long term, and who is responsible for each action. Agents contribute to that memory.

Claude is its primary model provider. Arca built its own harness, which picks a model for each task based on internal benchmarks, and owning it lets the firm switch providers. Agents analyze and propose. Trading and any money movement is handled by a human advisor. "All of our safety restrictions are fully deterministic," Singh said. Code the agent writes to query data is translated into a strict format with guardrails before it reaches the back end. A meta agent reviews advisor feedback and agent traces to suggest harness changes.

My take

Swap "wealth client" for "deal" or "patient" and this is how I would build an agent in HubSpot or GoHighLevel:

1. The CRM record is the memory. Store status, next step, open commitments and owner in fields the agent reads and writes, not in a chat history.
2. The agent drafts and suggests. Anything that moves money, signs, or messages a customer goes through an approval step written in plain logic.
3. Collect the corrections your team makes and review them weekly. That is your eval set.

Long running work is where agents save the most time, and also where they drift. Structure is what keeps them useful.