Agents Prompt People: Using Agents for Long-Horizon Financial Work

Aman Singh, lead AI engineer at the AI-native wealth management firm Arca, described the company's agent architecture at AI Engineer New York this month and in a follow-up conversation with Latent Space. Arca recently raised $48.5 million in a round led by General Catalyst, with Index Ventures and Venrock also participating. The defining feature of Arca's setup is an inversion of the usual interaction pattern: rather than staff queuing up prompts for an assistant, the agents themselves surface work and prompt the humans. Singh's phrasing captures it — agents prompt people more than people prompt agents.
The reason is the shape of financial work. Singh described agent tasks at Arca as long-horizon but not long-running, with a single task potentially stretching across ten days. In practice that means a stop-start cadence: many actors, many decision points and many handoffs along the way rather than one continuous model session. That is a very different engineering problem from a chat agent that answers in seconds. Anything spanning days needs durable state, clear ownership of each pending action and a way to resume without losing context.
Arca addressed this with an internal memory system that persists per client. It tracks what has already happened, what is being monitored next, what must occur in the coming days, what is being watched over the longer term, and which individual is responsible for which action. Agents contribute to that memory as they work. This is a useful design signal for teams building similar systems: the memory is not just a conversation log but an operational ledger of commitments and owners, closer to a case-management system with an LLM interface than to a chatbot with a history buffer. In practice, a ten-day stop-start workflow also implies idempotent tool calls, explicit approval states and audit trails, since a task may be picked up by different people at each handoff.
On models, Arca draws from frontier labs including Anthropic and OpenAI, with Claude as its primary provider. The company built its own harness, which automatically selects the best model for a given task based on Arca's internal benchmarks. Owning the harness instead of adopting a managed agent framework gives Arca the freedom to swap providers when needed. This is a familiar build-versus-buy trade-off: a managed harness offers speed and maintenance offload, while a custom one costs engineering time but buys routing flexibility, provider independence and the ability to encode firm-specific evaluations. For a regulated business where model choice may shift with pricing, latency or capability, that optionality can be worth the maintenance burden.
Safety is where the architecture is most distinctive. Agents do much of the background work, but their role is confined to analysis and proposing solutions. Trading and any other movement of money is handled by a human advisor. Singh said all of Arca's safety restrictions are fully deterministic and take decision-making off the agent's hands. Concretely, agents can execute some tools and tasks autonomously, but not ones that put client assets at risk. The word deterministic matters here: a probabilistic model can be argued into a bad action, a hard-coded rule cannot, so gating irreversible, high-stakes operations behind non-model logic is a sensible pattern for any agent that touches money.
Arca also runs verification checks against known AI security issues such as prompt injection. The agent itself is not sandboxed, but sandboxes are used for computation. Sensitive data handling goes a step further: code the agent writes is translated into a strict format the firm expects, the original code is discarded, and guardrails are applied to that format before it reaches the backend. This lets agents query data with code — aggregating, reshaping and working with it to produce outputs such as a financial plan — without granting raw execution against production systems. A likely trade-off is expressiveness: a strict intermediate format will reject or constrain some queries a free-form script could run, in exchange for a much smaller attack surface.
Finally, Arca closes the loop on its own harness. Feedback from human advisors and traces from agent runs are fed to a meta-agent that analyzes the data and suggests changes to the harness. This is essentially turning day-to-day operations into an evaluation and improvement pipeline, and it fits the company's broader effort to convert operational knowledge into agent context.
Why it matters: Teams building agents for finance, insurance or other regulated workflows face the same long-horizon, multi-actor problem Arca describes, and its answers — per-client persistent memory, a self-owned routing harness, deterministic gates on irreversible actions, and a sanitized code-to-query path — are a concrete template. The split of labor, with agents analyzing and proposing while humans approve trades and money movement, is likely to be the compliance baseline rather than a temporary compromise, in this reading. Expect more firms to invest in their own harnesses and evaluation loops precisely so model providers remain swappable.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Arca's head of AI discusses persistent memory, the benefits of a custom harness, and agentic workflows that span days.