Arca uses agents to analyze client finances, remember unfinished work, and bring advisors into a process when their judgment is needed. Aman Singh, who leads its AI platform team, describes a wealth-management firm built around that division of labor.
Ten days of work need shared memory
A client emails about a $100,000 tax bill. Agents examine the client’s assets and propose how to raise the money. The process spans roughly ten days, with pauses, people, and approvals between steps. It is long-horizon work, not one model thinking continuously for ten days.
An email reply can return to the agent following that thread. A separate phone call can wake a new agent. Persistent client memory tells it what has happened, who owns each action, and what needs attention next. Daily reflection updates that picture.
“Agents prompt people more than people prompt agents,” Singh says. The point is to notice when the next person is needed and give them enough context to act.
Humans own the consequential decisions
Agents propose trades. Human experts decide whether those trades are appropriate and execute them. Money movement, asset liquidation, and direct client communication require human involvement. Clients receive service through their advisor.
Arca has custodian APIs, but permission is constrained outside the model. Singh describes deterministic restrictions alongside additional language-model checks:
- Fast models inspect database writes.
- Incoming documents and messages are checked for prompt injection, where external content tries to redirect an agent.
- Outgoing data receives checks for personal information and attempted leakage.
- Generated financial plans are checked for account numbers and language promising returns.
Every agent action and human edit creates a change log. The same record supports audits and tells agents that something changed.
Keep the business knowledge when models change
Arca builds its own harness, the software that runs models and tools. Internal tests determine model choices, with heavier models for tax analysis and cheaper tiers for background work. Advisors do not choose the models themselves.
Singh wants Arca to retain control over memory and operational knowledge while switching model suppliers when needed. Because Arca operates the firm, it can also standardize the work its agents coordinate.
Experts maintain a file-based knowledge base in a Knowledge tab. That guidance differs from client memory: it describes how the firm works. Decision models nudge agents toward current financial facts or web search when needed. Singh’s example is a model confidently using outdated tax brackets.
Improvements still pass through an engineer
Staff explanations for edits and cancellations are paired with agent traces. A nightly meta-agent reviews that evidence and suggests fixes to tools, prompts, or missing knowledge. Engineers approve prompt changes before release. Singh says agents left alone tend to let one narrow correction consume the whole prompt.
The closing example is similarly concrete. Agents reportedly found additional fees and a non-fiduciary advisor while reviewing a prospective client’s documents. It is a story about scrutiny and service, with no measured investment-return improvement supplied.
Caveat: These are Singh’s descriptions of Arca’s processes, without independent safety or performance results. Autonomous trading is explicitly outside the current system.
Full conversation with Richard MacManus on Latent Space, October 31, 2026. The company discussed is Arca wealth management.