Wealth manager Arca lets agents do the groundwork, but people make the decisions and carry out money transactions.
Why this is useful to know
Arca is a wealth manager that uses AI as the foundation of its work. Agents do most of the groundwork: analysing, calculating and making proposals. The adviser makes the final decision, and trading or moving money is always done by a person. Aman Singh, head of AI at the company, sums it up like this: agents direct people more than the other way round. That is the point to remember. A task can take ten days, with many handovers between people. That is why Arca keeps a memory per client, so nothing gets lost. The lesson lies mainly in the division: what may the machine do on its own, and what stays with a person?
When you need this
Do you recognise work that takes a long time, passes through many hands and keeps hitting a decision point? Then this is the setup that works. Arca's principle: the agent may analyse and make proposals, but not execute. Anything that touches money or puts a client at risk goes through a person. Set that boundary in rules that do not depend on the model. Do the same in your own processes: decide per step what the AI may do itself, what needs approval and what never happens automatically. Also pay attention to handovers. An agent that works over several days needs a log of what has already happened and who has to do what. Without it, everyone loses track.
How it works under the hood
Arca builds its own harness: the layer that drives the model. It picks the model per task based on internal benchmarks and routes the work. Because the harness is theirs, they can switch model provider without rewriting their workflows. The memory per client is an internal system that the agents contribute to themselves. The agent itself does not run in a sandbox, but the code it writes does. The agent's output is converted into a strict format, the generated code is thrown away, and a guardrail sits on that format before it reaches the backend. A meta-agent analyses feedback from advisers and agent traces and proposes changes to the harness. That calls for your own evals and maintenance. Prompt injection is tested with verification checks, but that does not remove the risk.
This news touches chapters 11, 21 and 23 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a publication by Latent Space: https://latent.space/p/arca/
