Arca’s AI agents guide human advisors in long-horizon financial workflows
Arca, an AI-native wealth management firm, uses agentic systems with persistent memory and custom harnesses to support multi-day financial tasks, while keeping final decisions with human advisors.
1 source · cross-referenced
- Arca employs AI agents to handle long-horizon financial workflows that can span up to ten days, with tasks involving multiple decision points and handoffs.
- The company uses a custom harness to route tasks to frontier models like Anthropic’s Claude and OpenAI’s models, based on internal benchmarks.
- Agents prompt human advisors more than the reverse, with safety guardrails ensuring deterministic restrictions on agent actions, particularly for trading or money movement.
- Arca’s agent infrastructure includes persistent memory for client tracking, sandboxes for secure code execution, and a meta-agent to refine the harness using advisor feedback.
Arca, an AI-native wealth management firm, has implemented agentic workflows that support long-horizon financial tasks, which can take up to ten days to complete. These tasks are characterized by multiple decision points, handoffs, and actors, requiring persistent tracking of progress and responsibilities. Aman Singh, Arca’s head of AI, described the system as one where “agents prompt people more than people prompt agents,” emphasizing the role of AI in guiding human advisors rather than replacing them.
To manage these workflows, Arca has developed an internal memory system that tracks client-specific information, including past actions, immediate priorities, and long-term tracking needs. This system ensures continuity across multi-day tasks and assigns responsibility for specific actions. Singh noted that the memory system is continuously updated by agents, providing a foundation for both short-term and long-term planning.
The company uses frontier models from Anthropic and OpenAI, with Anthropic’s Claude serving as the primary provider. Arca has built a custom harness to automatically route tasks to the most suitable model based on the company’s internal benchmarks. Owning the harness allows Arca to switch model providers as needed, without being locked into a single vendor. Singh highlighted the importance of this flexibility in a rapidly evolving AI landscape.
Safety and compliance are central to Arca’s agentic systems. While agents handle analysis and propose solutions, trading and money movement are reserved for human advisors. Singh emphasized that safety restrictions are deterministic, removing decision-making authority from agents for high-risk actions. The company also implements verification checks for known AI security issues, such as prompt injections, and uses sandboxes for secure code execution.
Arca’s agent infrastructure includes additional security measures for handling sensitive data. The company has developed systems to translate agent-generated code into a strict format, discard the original code, and install guardrails on the translated format before it reaches backend systems. This approach allows agents to query and manipulate data—such as aggregating or transforming it to produce financial plans—while minimizing risk.
The company also leverages feedback from human advisors and agent traces to improve its harness. A meta-agent analyzes this data to suggest changes, effectively turning operational knowledge into evaluative improvements. This iterative process ensures that the harness evolves alongside the needs of the business and the capabilities of the models.
Singh contrasted Arca’s AI-native approach with firms that merely use AI tools, emphasizing the importance of owning the entire agent infrastructure. This ownership enables Arca to tailor its systems to the specific demands of financial services, where auditability, compliance, and safety are paramount.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74