Arca deploys tiered models and deterministic guardrails for financial agents
Arca has built a model-agnostic agent harness that routes financial tasks to specific model tiers while enforcing strict human-in-the-loop controls. The system prevents autonomous execution by requiring manual approval for all trades and money movements.
Arca’s agent harness is designed to be model-agnostic, currently utilizing frontier models from Anthropic as the primary engine and OpenAI as a secondary option, with a history of using Gemini. The system routes tasks to different model tiers based on complexity; heavy analytical tasks such as tax analysis are assigned to high-capability models like Opus or heavy GPT variants, while background activities are handled by lower-tier models. To maintain control, Arca argues that owning the harness and intelligence layer is critical for business control, avoiding dependency on external managed agent products.
Safety is a central feature of the architecture. Agents do not execute trades or move money autonomously; instead, they perform analysis and propose plans that require human-in-the-loop approval for execution via custodian APIs. Deterministic checks and LLM-based verification are applied to every data write to ensure safety. The system uses sandboxes for computation, where agents generate code to query data and produce artifacts such as PDF financial plans. These artifacts undergo verification for personally identifiable information and compliance issues, including checks for prohibited promises of returns.
A component called the 'decision model,' which includes Jev, nudges agents to use verified current facts. This mechanism prevents errors such as using outdated 2025 tax brackets. Arca noted that its agents prompt people more than people prompt agents, reflecting a design philosophy focused on proactive assistance within strict safety boundaries.