Arca, an AI-native wealth management firm that just raised $64m, says agents replace the work around its advisers, not the advisers themselves.
Aman Singh says Arca moved from Gemini to OpenAI to Anthropic models, with Claude as primary, after Llama proved too slow and costly despite benchmarks.
Arca keeps agents from moving money or liquidating assets automatically, with human guardrails on trades and client communication and no autonomous trading today.
How Arca's agents run a wealth firm behind human advisers
Arca is a wealth management firm run by agents that work behind human advisers. Financial tasks stretch over days and carry real risk, so the agents need persistent client memory, deterministic guardrails and human sign-off on trades and client messages. Singh treats that in-house intelligence and memory, rather than any single model or vendor harness, as the firm's lasting advantage.
Takeaways
AI replaces work around advisers, not advisers. Arca, which recently raised $64m, is the wealth management firm itself and not a SaaS vendor. Clients from early-retiring tech executives to first-time home buyers deal with a human adviser while agents handle back-office work and monitor finances around the clock.
Model choice follows internal benchmarks, not hype. Claude is the primary model after moves from Gemini to OpenAI to Anthropic. Llama failed as too slow and costly despite strong external benchmarks. Arca's own tiered harness routes tax plans to heavy models and background work to lighter ones.
Persistent memory lets agents resume work days later. Memory systems track each client's state and who owns each task. A daily dreaming process reflects on the client's situation to decide when to pull in a person. Fresh signals, such as a phone call, wake new agents that read this memory.
Safety is deterministic, with humans approving trades. Agents never move money or message clients on their own. Every data-writing tool call gets LLM verification, edges are checked for prompt injection and PII, and a consensus of small models blocks exfiltration. Change logs make agent and human actions auditable.
Verifiers check every artifact before clients see it. Agents offload computation to sandboxes that convert code output into a strict format with guardrails. Deterministic and fast-model verifiers scan client PDFs for raw account numbers or promised returns, while decision models catch stale facts like confidently used 2025 tax brackets.
Owning intelligence and memory is the moat. Singh is bullish on Anthropic and OpenAI models but wary of depending on Anthropic's Managed Agents. He calls Anthropic's push into financial services validating, and defines an AI-native firm as one where agents prompt people rather than people prompting agents.
Staff edits become evals and harness changes. Feedback is a first-class product feature, stored with the agent trace that produced it. A nightly meta agent reviews real traces and proposes harness changes that a human engineer approves. Agents cite a curated knowledge base staff can edit in the product.
A ten-day tax bill, step by step
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Agents analyse where the client's money is allocated to find the best way to raise $100,000.
The task spans ten days, so agents pick it up again from client memory as new signals arrive.
Agents draft a plan for the adviser to review, since trades and client communication require human approval.
Experts decide whether to act, and execution runs through Arca's deep API integration with its custodian.