Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Summary
Arca is building an AI-native wealth management firm, focusing on proactive, 24/7 service that aims to replace the 'work around the advisor' rather than replacing the advisor themselves. The firm's core technology is a custom, model-agnostic harness designed to run complex, long-horizon agentic workflows (spanning days). Key architectural components include persistent memory systems for tracking client state, deterministic safety guardrails, and sandboxes for computation, ensuring compliance and auditability by default.
Key takeaways
-
AI's Role: Proactive Service, Not Replacement
Arca's AI is designed to provide a level of service that is proactive and responsive, handling the 'back office stuff' of finance. The goal is for the client to experience the care of a human advisor, while the AI manages the complex, continuous monitoring of market and life events impacting finances.
-
Custom, Model-Agnostic Architecture
6:41
Arca utilizes its own internally built harness, which is model-agnostic, allowing flexibility across frontier models (e.g., Anthropic, OpenAI). Model selection is guided by internal benchmarks based on task complexity, assigning tasks to different tiers of models (e.g., 'heaviest tier' for tax analysis, lower tiers for background activity).
-
Long-Horizon Workflow Management
12:00
Agentic workflows can span days (e.g., adjusting a $100,000 tax bill). Persistent memory systems track the client's state, ownership of tasks, and historical context, allowing fresh agents to 'wake up' and contribute to ongoing processes, even if the initial signal was unrelated.
-
Safety and Compliance by Design
18:22
Safety is enforced by fully deterministic guardrails that take decision-making off the agent's hands. Verification layers include LLM-based checks on every data write, edge detection for prompt injection, and deterministic checks for PII during external data transmission.
Technical details
-
Agentic Workflow and Harness
401s
Arca operates a custom, internal harness that is model-agnostic. This harness manages tasks by assigning them to model tiers based on complexity (e.g., 'heaviest tier' for tax analysis). The system abstracts model selection from the human advisor, making decisions internally.
-
Data Integrity and Sandboxing
1505s
Sandboxes are used to offload computation, allowing agents to query and aggregate data. The code written within the sandbox is translated into a strict format before reaching the backend. Artifacts produced in the sandbox undergo multiple verifications: deterministic checks (e.g., preventing raw account numbers) and fast LLM checks (e.g., preventing non-compliant promises of returns).
-
Knowledge and Feedback Loops
The system incorporates a 'Knowledge' tab containing the full operational context, which is curated by human experts. The platform facilitates a feedback loop where staff edits or cancellations are captured as 'evals.' These evals are piped into a data store, allowing a meta-agent to run nightly processes that suggest improvements to the harness, tools, or knowledge base.
-
Decision Models for Financial Rigor
To prevent models from confidently using incorrect or outdated facts (e.g., 2025 tax brackets), Arca uses decision models within its harness. These models detect when a financial recommendation is made and nudge the model to use the knowledge base of verified facts or web search results.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.