Arca keeps financial agents on 10-day jobs but keeps money movement human

Financial firm Arca lets agents coordinate client work that can span up to 10 days. The model does not run continuously. The system saves the case state, waits for an email, phone call or human decision, then resumes where the work stopped.

Client memory holds a 10-day job together

Aman Singh, who leads Arca's AI platform, described workflows with many handoffs and decisions to Latent Space. Per-client memory records what has happened, what comes next, who owns the next action and what needs long-term monitoring. A new signal can therefore wake a fresh agent that loads the state of an active case.

Arca uses models from Anthropic and OpenAI. Its custom harness routes work according to internal benchmarks and task complexity. The firm names Claude as its primary provider, but keeps the orchestration layer so it can change models without handing its operational memory to one vendor.

The agent owns coordination while the adviser owns the decision

The important shift concerns who prompts whom. The agent keeps watching the goal and brings in an adviser when approval or intervention is required. A person does not have to reconstruct the case after every pause and manually issue the next prompt.

For teams building a similar system, the real product is the workflow state model. Prompts and model selection are only components. The decisive layer is a durable record of responsibilities, events and next actions that survives any single model run.

Money movement stays behind a hard human boundary

Arca says agents automate analysis and proposed solutions, not trades or other money movement. In a $100,000 tax-bill example, an agent can prepare an asset liquidation plan, but a person executes the trade. Fast models also check data writes, while incoming material is screened for prompt injection.

That is a sensible architecture, but the evidence is still largely the company's own account. Public data does not show error rates, adviser intervention rates or how many proposals the verification layer stops. Without those figures, it is impossible to tell whether memory cuts costs or merely moves coordination into more complex software.

Audit trails and intervention rates will reveal the real cost

The strongest signals will come from routine operations: how many 10-day cases finish without lost context, how often a person corrects the plan and how frequently a safety check blocks a bad action. Arca says it logs changes by both people and agents, so it has the trail required to measure those outcomes.

A second test will be a model switch. A custom harness earns its keep if Arca can actually change providers without breaking workflows, memory or controls. That is where model independence must work beyond the presentation.

Lilith's verdict

Arca put the agent on the case for 10 days but left the vault with the adviser. That is useful financial autonomy: long memory, a short leash and a human hand on the money.