Source time Edition date
ToolsLatent Space
Agents Prompt People: Using Agents for Long-Horizon Financial Work
智能体提示人们：用智能体开展长期金融工作
Arca's head of AI makes the case that long-horizon financial agents hinge on persistent memory and a custom harness, so teams must weigh that build cost against off-the-shelf simplicity.
Judgment
AI engineering leads at wealth-management firms must decide whether to own persistent client memory and a custom model-routing harness, as Arca does, or accept managed-harness convenience with less provider flexibility; Arca keeps trading decisions with human advisors.
Basis
Aman Singh, lead AI engineer at Arca, told Latent Space that agent tasks can take up to ten days and are stop-start, with many handoffs, so Arca maintains persistence for every client—tracking what happened, next actions, ownership, and long-term items. Arca uses frontier models, primarily Claude, and built its own harness that selects the best model per task using internal benchmarks; owning it preserves the option to change providers. Safety is deterministic: agents analyze and propose, but trading and money movement stay with a human advisor. Arca also translates agent-written code into a strict format and discards the original before backend access. Arca raised $48.5 Million led by General Catalyst.
Unclear The source does not quantify the cost, latency, or failure rates of Arca's custom harness versus a managed alternative, nor how much advisor time the approvals consume.
Next Ask Arca or Singh for a concrete before/after metric—such as model-routing accuracy or advisor approval time—when the custom harness is compared with a managed harness.