Arca gives AI agents finance tasks lasting up to 10 days, but leaves the trades to humans

Arca is an "AI-native" wealth management company, and AI agents do almost all of its background work. A single workflow can take up to ten days. That is not continuous computation but a long calendar-time process with pauses, many participants and handoffs. Aman Singh, Arca's lead AI engineer, said at the AI Engineer conference in New York that live advisers still have the final say. He put it this way: "agents prompt humans more often than humans prompt agents." Arca recently raised $48.5 million in a round led by General Catalyst, with participation from Index Ventures and Venrock.

Long horizon instead of long runtime

Singh distinguishes between two things: tasks in finance are long in calendar terms, but the agent does not spin non-stop. In his words, the work moves in bursts, with many actors, decision points and handoffs between steps. One example discussed on the podcast is a $100,000 tax bill, where agents prepare the analysis and a human gives the approvals.

To keep an agent from losing the thread over such a stretch, the company built its own memory system. For each client, it records what has already happened, what is being tracked next, what needs to be done in the next couple of days and what matters over the long term. It also records which specific person is responsible for which action. Agents don't just read this memory; they update it themselves.

An in-house harness and model selection

Arca uses models from frontier labs, primarily Anthropic and OpenAI, and Claude remains its main provider. An engineer does not pick the model for a given task by hand. The company's own harness handles that, relying on its internal benchmarks.

Singh explains that owning the harness is about flexibility. A managed agent harness from a vendor ties a team to that vendor's models, while in-house infrastructure lets it switch providers if it ever needs to.

There is also a self-improvement loop. Feedback from advisers and traces of agent work are collected together, and a "meta-agent" analyzes them and proposes changes to the harness. The podcast also discusses turning employee feedback into evals, which are regular quality checks.

Deterministic constraints and sandboxes

In finance, security is the price of entry. The agent's role at Arca is limited to analysis and proposing decisions. Trading and any movement of money remain with the human adviser. "All of our safety constraints are fully deterministic," Singh says. "They take decision-making away from the agent." The agent calls some tools on its own, but anything that carries risk for the client is off limits to it. The company also tests for known vulnerabilities of AI systems, including prompt injections.

The agent itself does not run in a sandbox, but it does its computation there. An additional layer of protection governs data access. Code that the agent wrote in the sandbox is converted into a strict, expected format, and the original code is discarded. Constraints are applied to that format before the request reaches the backend. Agents can query data with code, and aggregate and transform it as they see fit, for example to build a financial plan.

What you can take away

Arca's experience is a practical template for industries where mistakes are costly. Memory is better built around the client or case than around the conversation: that way a multi-day process survives pauses and handoffs between people. Permissions are better enforced by rules in code than by instructions in a prompt. Converting agent-written code into a verifiable intermediate format removes some of the risks associated with sandboxes.

There are caveats, too. This piece is a retelling of an interview and a presentation, not an independent audit: there is no data on agent errors, time saved by advisers or the number of clients. The podcast topics also include compliance and auditability "by default," proactive client outreach and finding hidden fees, but the text gives no details. The overall picture is clear: in long-running work with money, the agent prepares a decision and reminds the human of the next step, but the human still makes the decision.