Key points
- Arca, an AI-native wealth management firm, closed a $48.5 million funding round with General Catalyst as lead, plus Index Ventures and Venrock.
- Agent tasks at the firm can stretch to ten days, so Arca built a per-client memory system that tracks status, next steps and who owns each action.
- Agents only analyze and propose; a human advisor executes trades and any movement of money, under fixed rules the agent cannot override.
- Claude is the primary model, and Arca's own harness picks a model per task using internal benchmarks.
Why it matters for marketers: Arca shows a working pattern for agent workflows with many handoffs: persistent memory, owned routing, and hard-coded limits that keep humans on the consequential decisions.
Arca, a wealth management company that calls itself AI-native, has put agents at the center of its background work while keeping licensed advisors in charge of outcomes. Lead AI engineer Aman Singh described the setup at AI Engineer New York. In his words, "agents prompt people more than people prompt agents." The company also closed a $48.5 million round led by General Catalyst.
The work is long-horizon rather than long-running: a single task can span up to ten days, with many handoffs and decision points. To cope, Arca built a memory system that keeps a record for each client, covering what has happened, what comes next in the short and long term, and which person is responsible for each action.
Safety rules are deterministic, so the agent does not decide where its limits sit. Agents can analyze and propose, but trades and money movement go to a human advisor. Arca also runs checks for known problems such as prompt injection, and it converts code written in a sandbox into a strict format with guardrails before it touches back-end data.
For marketers and content teams, the lesson is structural. Campaigns, approvals and publishing also involve stop-start work across many people, and the same pieces apply: durable memory, clear ownership of each step, and fixed boundaries around anything risky such as spending or publishing.
Arca also owns its harness, which routes tasks to models based on internal benchmarks and leaves room to switch providers. It feeds advisor feedback and agent traces to a meta-agent that suggests harness changes. Teams building their own agent stacks may want to watch that feedback loop.
Sources
This brief is our own summary. Read the full reporting at the source.
