Arca's approach to long-horizon financial agents and safety

At the recent AI Engineer New York conference, Aman Singh, lead AI engineer at wealth management firm Arca, detailed how his team builds agents for complex, multi-day financial workflows. The company recently secured $48.5 million in funding led by General Catalyst, with support from Index Ventures and Venrock, to scale this AI-native infrastructure.

What happened

Singh explained that while AI agents handle the majority of background analysis and preparation at Arca, they do not operate in isolation. The firm has implemented strict guardrails ensuring that human advisors retain final decision-making authority, particularly for high-stakes actions like trading or moving money. This structure flips the traditional interaction model, where humans typically prompt bots. Instead, the agents analyze data, propose solutions, and then prompt the human advisors to review and approve specific steps.

The nature of financial planning requires what Singh calls "long-horizon" work. Unlike simple query-response tasks, agent workflows at Arca can span up to ten days. These processes are stop-start, involving multiple actors, decision points, and handoffs between automated systems and human staff. To manage this complexity, Arca developed an internal memory system that maintains persistence for every client. This system tracks what has occurred, what is pending in the immediate future, and what is scheduled for the long term, assigning responsibility for each action to specific individuals or automated processes.

How it works

Arca relies on models from major providers, using Claude as its primary engine, but it does not depend on a single vendor’s managed tools. The engineering team built a custom harness that automatically routes tasks to the most suitable model based on internal benchmarks. Owning this orchestration layer gives Arca the flexibility to switch providers if performance or cost dynamics change, without rewriting their core application logic.

Security and safety are handled through a combination of deterministic rules and isolated computation environments. The agents themselves are not fully sandboxed, but any code they generate for data processing runs inside secure sandboxes. To prevent security risks like prompt injections or unauthorized data access, Arca translates agent-written code into a strict, expected format. The original code is discarded, and only the sanitized format is allowed to interact with the backend. This allows agents to aggregate and munge data for financial plans without exposing the core database to arbitrary executable code.

Furthermore, all safety restrictions are fully deterministic, removing ambiguity from critical decisions. If a task involves risk to a client, such as executing a trade, the agent is technically incapable of performing it alone. It can only prepare the analysis and present it for human approval. A "meta-agent" also analyzes feedback from human advisors and agent traces to suggest improvements to the harness, creating a continuous loop of operational refinement.

Key details

• Agent tasks at Arca can take up to ten days to complete due to the stop-start nature of financial planning.
• The company raised $48.5 million in a round led by General Catalyst, with participation from Index Ventures and Venrock.
• Arca uses a custom-built harness to route tasks to the best model, primarily relying on Claude from Anthropic.
• An internal memory system provides persistence for every client, tracking immediate and long-term actions across multiple handoffs.
• Safety restrictions are fully deterministic, ensuring agents cannot execute trades or move money without human approval.
• Code generated by agents is translated into a strict format before reaching the backend, discarding the original code to prevent injection attacks.

Why it matters

For engineers building agentic systems, Arca’s architecture highlights the importance of persistence in long-running workflows. Most current agent frameworks assume short-lived interactions, but real-world enterprise tasks often require state management over days or weeks. By building a dedicated memory system that tracks context, responsibilities, and timelines, Arca avoids the context loss that plagues many multi-step agent implementations. This approach ensures that when an agent resumes work after a human review or a waiting period, it retains full awareness of the client’s status.

The separation of analysis from execution is another critical lesson. By limiting agents to proposing solutions and enforcing deterministic guardrails for high-risk actions, Arca mitigates the unpredictability inherent in large language models. This design pattern allows firms to leverage the speed and breadth of AI for data processing while maintaining the accountability and safety required in regulated industries. It demonstrates that trust in AI systems comes not from making the model perfect, but from constraining its ability to cause harm.

What you can do

• Implement a persistent memory layer for your agents to track state across multi-day workflows and handoffs.
• Build a custom routing harness to benchmark and select the best model for specific subtasks rather than relying on a single provider.
• Enforce deterministic guardrails for any action that involves financial risk, data deletion, or external side effects.
• Sanitize agent-generated code by translating it into a strict, limited format before allowing it to access your backend data.
• Use human feedback and agent traces to train a meta-agent that suggests improvements to your orchestration logic.
• Design workflows where agents prompt humans for approval at key decision points, rather than expecting humans to initiate every interaction.