Speaking with Latent Space at AI Engineer New York, Aman Singh, lead AI Engineer at wealth management firm Arca, described how the company uses AI agents across long-horizon work. Singh said agent tasks at Arca can take up to ten days and are 'very stop-start,' with 'many actors in between, many decision points, many handoffs.' To manage this, Arca built an internal memory system that tracks, per client, what has happened, what is being tracked next, what must happen immediately, and which individual is responsible for each action. According to the interview, agents perform background work but their role is limited to analysis and proposing solutions; trading and money movement are handled by human advisors. Singh said Arca's safety restrictions are 'fully deterministic' and take decision-making 'off the agent's hands.' The firm uses models from Anthropic and OpenAI, with Claude as its primary provider, and has built its own harness that selects models based on internal benchmarks and allows switching providers. Arca runs verification checks for issues such as prompt injection, and uses sandboxes for computation with added measures for sensitive data. A 'meta-agent' analyzes advisor feedback and agent traces to suggest harness improvements. Latent Space reported Arca recently raised $48.5 million in a round led by General Catalyst, with Index Ventures and Venrock participating.
- Agent tasks at Arca can take up to ten days and involve many handoffs and decision points.
- Trading and money movement stay with human advisors; agents only analyze and propose.
- Safety restrictions are described as 'fully deterministic,' removing certain decisions from the agent.
- Arca uses Claude (Anthropic) and OpenAI models via a custom in-house harness.
- Arca raised $48.5M led by General Catalyst, with Index Ventures and Venrock.
What it means for you
This is a real-world example of a pattern worth copying, not a product you can buy. A firm doing high-stakes work (wealth management) lets agents do the grinding background work but keeps every decision that touches money with a human. The useful idea here is their memory system: for each client they track what happened, what's next, what's urgent, and who owns each action — that's a process discipline, and the AI is bolted on top of it, not the other way around.
Try this
Pick one long, stop-start process your team runs (onboarding, a claim, a multi-step client request) and write down, for one live case, the same four things Arca tracks: what's happened, what's next, what's urgent, and who owns each step. If you can't answer those cleanly today, you've found the real problem — and no tool fixes it for you.
Who should care
Owners and operations leads running regulated or high-consequence work with many handoffs — finance, insurance, legal, healthcare — and anyone weighing whether to put agents on multi-day workflows.
Skip this if
Your work is short, self-contained tasks, or you're not building custom AI systems — this is one firm's in-house engineering approach, not something you can plug in off the shelf.
Sources: Latent Space — read the original