Agents Prompt People: Using Agents for Long-Horizon Financial Work
- ID
- 41574
- Status
- summarized
- Published
- 01 Nov 2026, 12:05 AM
- Fetched
- 01 Nov 2026, 1:01 AM
- Provider
- Latent Space
- Category
- developer-ai
- Original URL
- https://latent.space/p/arca/
- Source URL
- https://www.latent.space/feed
Summary
- Score
- 6.5
- Created
- 01 Nov 2026, 1:01 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_startup_founders
What happened
At AI Engineer New York, Latent Space spoke with Aman Singh, lead AI Engineer at Arca, an "AI-native" wealth management firm that recently raised $48.5M led by General Catalyst with Index Ventures and Venrock. Singh said agent tasks at Arca can span up to ten days but are "not long-running" — they are stop-start, with many actors, decision points and handoffs, so the firm built an internal per-client memory system tracking what has happened, what is due in the next few days, and what is tracked long term. Human advisors still make the final call; Singh's framing is that "agents prompt people more than people prompt agents." The published text is a partial transcript.
Why it matters
The useful design claim is the distinction between long-running and long-horizon: if your agent's work spans days with human handoffs, a chat transcript is the wrong state store. Arca's approach — a per-client persistent record of past events, near-term next actions and long-term tracking, with advisors as the approval gate — is a concrete pattern to copy when you scope agent memory and where the human sign-off sits. No Malaysia-specific angle appears in this excerpt.
Discussion angle
How would you model Arca's stop-start state: an append-only event log per client, a rolling summary, or an explicit task queue with owners and due dates — and at which handoff points does a human approval gate actually earn its friction?