AI News
All stories
Reported

Arca says its agents run tasks up to ten days, with humans approving trades and payments

Arca's head of AI says its agents handle long-running wealth management work using a custom harness and memory system, while humans make final financial decisions.

AI in Practice · · 2 min read

Arca, a self-described AI-native wealth management firm, says its agents handle most of the background work while human advisors make the final financial decisions. A single agent task can run for up to ten days. Aman Singh, the firm's lead AI engineer, described this long-horizon work on the Latent Space podcast as "very stop-start," with many handoffs and decision points.

Arca keeps a persistent memory for each client, and its agents write to it. Singh said the memory records what has happened so far, what is being tracked for the next few days and over the long term, and which individual is responsible for each action.

Arca uses models from frontier labs, including Anthropic and OpenAI, with Claude as its primary provider. Its own harness, the software layer that wraps the models and routes work to them, picks a model for each task based on the company's internal benchmarks. Singh said owning the harness, instead of using a managed agent harness, gives Arca the flexibility to change providers if needed.

On safety, Singh said the safety restrictions are fully deterministic: "All of our safety restrictions are fully deterministic," he said, adding that they "take the decision making off the agent's hands." In practice, agents can analyse and propose solutions, but trading and any other money movement stay with a human advisor. Arca also runs verification checks for known AI security issues such as prompt injection. The discussion did not list which tools fall on the agent's side of that line and which fall on the human side.

Access to client data gets extra security measures. Singh said agents write code to query client data, and that code is then translated into a strict format the company expects and discarded. Guardrails apply to that format before anything reaches back-end systems. The agent itself is not sandboxed, but it uses sandboxes for computation. Singh said this lets agents aggregate and reshape client data to produce outputs such as a financial plan, while the guardrails govern what reaches the systems behind them.

Arca also feeds advisor feedback and agent traces into a "meta-agent," which analyses that data and suggests changes to the harness. Latent Space reported that Arca recently raised $48.5 million in a round led by General Catalyst, with participation from Index Ventures and Venrock.

Why it matters

Firms running multi-day agent workflows in regulated settings can keep control by owning the harness, storing persistent state outside the model, and enforcing action limits in code rather than through prompts.

What you can do

Review which agent actions in your own workflows can be enforced with deterministic rules, and keep money-moving or client-facing steps behind human approval.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Latent Space

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works