VenturFlow Open-Sources Assay: Safety Layer for AI Agents in Finance
Key Takeaways
- ▸Assay provides four-layer validation for AI agent outputs in financial workflows, catching errors and policy violations before money moves
- ▸VenturFlow released the library as open-source to help the broader AI+finance community avoid duplicating safety infrastructure
- ▸Pre-built regulatory compliance rule packs (SEC, FINRA, MiFID II, OFAC) come with citations and effective dates, though they don't replace compliance counsel
Summary
VenturFlow has released Assay, an open-source Python library designed to validate and safeguard AI agent decisions before they execute financial transactions. Released under Apache 2.0, Assay serves as a critical safety and validation layer that sits between AI agents powered by LLMs (Claude, GPT, Gemini, etc.) and downstream financial actions like trades, wires, and regulatory filings.
The library addresses recurring safety challenges that VenturFlow encountered while building its own agentic AI platform for venture capital firms. Rather than forcing each organization to solve these agent-safety problems independently, VenturFlow extracted the validation layer and contributed it to the open-source community. The library provides four key validation boundaries: output validation with schema and regulatory rule packs, tool-call gating with typed arguments and approval thresholds, trajectory validation for session-wide constraints and loop detection, and entity resolution to ground ticker symbols, CUSIPs, and counterparties correctly.
Assay's built-in regulatory rule packs cover SEC 15c3-1, Regulation T, Volcker, FINRA 4210, MiFID II suitability, and OFAC sanctions—with explicit disclaimers that these are not substitutes for compliance counsel. All validation checks feed into an append-only audit log for complete transparency and auditability. The library follows a "bring your own data" (BYOD) model, ensuring sensitive firm data stays local unless users opt in to optional remote semantic-consistency checks via LLM providers like Anthropic or OpenAI.
- Assay maintains privacy through a BYOD model—firm data stays local unless explicitly configured to use remote LLM providers
