
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Latent Space: The AI Engineer Podcast
- 26m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
This episode of Latent Space features Aman Singh discussing how Arca operates an AI-native wealth management firm that recently raised sixty-four million dollars. The guest and host explore how large language models and custom internal harnesses power long-horizon financial workflows spanning up to ten days, freeing human advisors from back-office tasks. Rather than replacing advisors or letting clients prompt agents directly, Arca builds proactive background systems where agents prompt human staff and manage tasks like tax analysis and portfolio reviews. The discussion details technical guardrails, deterministic safety checks, sandbox usage for offloading computation, and how staff edits are systematically converted into model evaluations and harness improvements.
- Abstract model selection by routing tasks to different tiers based on complexity rather than locking into a single provider. 4:41
- Enforce strict human-in-the-loop guardrails for any actions involving money movement, asset liquidation, or client communication. 3:45
- Keep agent memory systems updated through daily reflection processes to handle long-horizon tasks lasting several days. 5:52
- Pipe product feedback and staff edits into a daily meta-agent process to suggest improvements to prompts and tools. 21:02
Chapters
- 0:00Arca’s AI‑Native Wealth Management Model
A brief intro to Arca’s non‑SaaS firm and its AI‑powered wealth management services.
- 2:44Model Selection and Harness Architecture
Discussion of the model tiers, internal benchmarks, and the decision to stay model‑agnostic.
- 3:57Client Success Example and Proactive Value
A real‑world case where agents uncovered hidden fees and improved client outcomes.
- 6:07Long‑Horizon Portfolio Management Workflow
Walkthrough of a ten‑day portfolio adjustment process and the agent’s responsibilities.
- 7:42Memory, Persistence, and Context Refresh
How Arca’s agents use internal memory and daily reflections to handle delayed signals and long tasks.
- 9:22Safety, Verification, and Compliance Controls
Deterministic guardrails, LLM verification, and audit logs ensure secure, compliant agent actions.
- 20:12Human‑in‑the‑Loop and Feedback Loops
Agents propose actions but humans approve, and feedback is logged to improve the harness.
- 22:41Knowledge Base and Operational Context
Agents cite a curated knowledge base that is continuously updated by experts.
Highlights
- 0:00
Agents Prompting Peopleinsight
The guest explains that Arca is not a traditional SaaS company, but an operational firm where agents proactively prompt people twenty-four seven rather than waiting for user prompts.
- 1:51
Background Automation for Advisorsstrategy
The role of AI at Arca is strictly to handle back-office work, allowing human advisors to deliver high-touch service driven by empathy and care.
- 3:41
Model Agnostic Harnessstrategy
The company relies on internal benchmarks rather than external claims, maintaining a model-agnostic approach across frontier labs like Anthropic and OpenAI.
- 5:52
Ten-Day Workflow Executionstory
Financial workflows like tax analysis and portfolio rebalancing can span ten days, with agents handling the deep analysis while humans retain execution authority.
- 9:32
Deterministic Safety Guardrailsinsight
All safety restrictions are fully deterministic and handle write operations via rapid LLM-based verification to prevent unauthorized data access or modifications.
- 11:18
Audibility by Defaultinsight
Every action taken by an agent or human produces a detailed change log, satisfying strict regulatory compliance while feeding the agentic loop.
- 14:17
Avoiding Managed Agentsdebate
The guest explains why Arca avoids vendor-managed agent frameworks, noting that owning the intelligence layer is critical to building a truly unique business.
- 21:31
Meta-Agent Prompt Optimizationstrategy
Arca runs a nightly meta-agent process over real-world traces and staff feedback to suggest improvements to harness tools and system prompts.
Quotes
“We are actually not a SaaS company. We are, um, we are the firm ourselves. Agents prompt people more than people prompt agents.”
Unnamed speaker · 0:00 “The role of AI at Arca is not to replace the advisor, it is to replace all of the work around the advisor.”
Unnamed speaker · 1:51 “We have every single action that an agent takes produces a change log of what changed, who changed it, from which part of the system and when.”
Unnamed speaker · 11:22 “Agents prompt people more than people prompt agents. And it is a proactive kind of service that is running for them 24/7.”
Unnamed speaker · 15:44
About this episode
Arca head of AI Aman Singh joins Latent Space head of editorial Richard MacManus at AI Engineer New York, to discuss how Arca is building an AI-native wealth management firm. Singh explains Arca's architecture, which includes a custom harness and persi...
Transcript
4,708 words · 2 speakers
Topics
- Arca
- sandbox
- Anthropic
- SaaS
- OpenAI
- Jev
- memory systems
- knowledge base
- Frontier Labs
- Gemini
Page updated · transcript and summary are machine-generated