Arca's AI agents work on client cases that last ten days

A client case at Arca can take ten days to move through the firm. The agent and several human advisors each take turns on it, step by step, with handoffs between them.

That is how Aman Singh described the work at Arca. Singh is lead AI engineer at the firm. He spoke to Richard MacManus on Latent Space on 31 October. Arca is an AI-native wealth management firm. It recently raised $48.5 million, Latent Space reported. General Catalyst led the round. Index Ventures and Venrock also joined.

How the memory works

Singh said Arca built a persistent memory for every client. He said it tracks what has happened so far. It tracks what the firm is watching for next. It tracks what has to happen in the next few days. It also tracks what the firm is holding for the long term. For each action, it names the person responsible. A case can run for ten days with many handoffs between the agent and human advisors.

Where the safety sits

Singh said all of the firm's safety limits are "fully deterministic." They take the choice out of the agent's hands. Any code the agent writes runs inside a sandbox. For sensitive data, Arca does more. It reads the model's code and turns it into a strict format the backend accepts. The model's own code is thrown away before anything reaches that backend.

Singh gave an example. On a $100,000 tax bill, the agent studies the case and proposes a path. A human advisor executes the trade.

The design cuts a class of risk out of the agent's reach. If the model is tricked into writing bad code, that code never runs on the real backend. The schema check throws it away. Audit and compliance come out of the same build. Each client case has a record of what the agent did and who approved it.

One harness, many models

Arca built its own agent harness, Singh said. It routes each task to the model that Arca's internal benchmarks say is best for that task. The firm uses frontier models from Anthropic and OpenAI, with Claude as the main provider. Owning the harness lets Arca swap providers without rebuilding the product. The firm has not published the benchmarks it uses.

Singh's own line in the Latent Space episode was that at Arca, agents prompt people more than people prompt agents. The claim is Arca's own, said about Arca's own product. No measured industry number is attached to it.

Source

• "Agents Prompt People: Using Agents for Long-Horizon Financial Work," Latent Space, Richard MacManus