ADK Financial Advisor Agent: A Case Study in AI Systems Safety Engineering

Introduction: A Praxis-Oriented Approach to AI Safety

This repository contains the source code for the Financial Advisor Agent, a sample project built using Google's Agent Development Kit (ADK). Its primary purpose extends beyond a simple demonstration of agentic capabilities; it serves as a live testbed for "The Google AI Systems Safety Engineer: A Praxis-Oriented Professional Development Framework."

Disclaimer: This project is intended for educational and research purposes in the field of AI systems safety. It is not intended for use in a production environment or for providing actual financial advice.

---

🆕 FinGuard: The Reference Architecture (ISO 42001)

We have introduced FinGuard, a new clean-slate reference architecture located in the finguard/ directory. FinGuard supersedes the previous financial-advisor demo for governance purposes, providing a production-ready blueprint for an Autonomous Agent System that complies with ISO/IEC 42001 and NIST AI RMF standards.

Key Features of FinGuard:

• Star Topology: A central CoordinatorAgent (Supervisor) manages specialized workers (Researcher, Quant, Compliance, Executor) to enforce separation of concerns.
• Policy-as-Code (OPA): Real-time transaction gating using Open Policy Agent (OPA) and Rego policies (finguard/policies/trade.rego), replacing Python-based mocks.
• Semantic Guardrails: A Semantic Drift Detector using Vertex AI Embeddings to prevent "Vaporwork" (infinite semantic loops) and enforce progress.
• Zero Standing Privileges (ZSP): Strict identity isolation where only the Executor agent possesses the IAM credentials to interact with the Brokerage API.
• Sandboxed Execution: Quant analysis runs in a simulated environment (designed for Cloud Run/gVisor) to prevent tool misuse and side-channel attacks.

👉 View the FinGuard Architecture & Structure

---

Legacy Architecture: VACP Integration

The original financial-advisor agent has been refactored to implement a robust safety architecture based on ISO/IEC 42001:2023 standards and a Verifiable Agentic Control Plane (VACP).

1. Zero Standing Privileges (ZSP) Architecture

We have migrated from static keys to a GCP-Native Zero Standing Privileges model.

• No Long-Lived Secrets: The application container has zero intrinsic permissions to trade.
• Identity Trade (MIM): The agent must perform an "Identity Trade" via the MIMService (Machine Identity Management) to exchange its low-privilege Workload Identity for a short-lived, high-privilege JIT token.

2. Verifiable Agentic Control Plane (VACP)

The VACP replaces the traditional "human-in-the-loop" with a "governance-in-the-loop" architecture, featuring:

• AgentGuard: CMDP-based runtime verification.
• Janus Shadow-Monitor: Continuous internal red-teaming.
• System 4 Derivative Estimator: Predictive risk estimation.

3. STPA-Driven Guardrails

Following a formal hazard analysis (see STPA_ANALYSIS.md), the VACP enforces hard safety constraints via Exponential Control Barrier Functions (ECBF).

Getting Started

Prerequisites

• Python 3.11+
• Poetry (for dependency management)
• Google Cloud Project (with Vertex AI enabled)
• OPA Binary (required for FinGuard policy evaluation)

1. Installation

Clone the repository and install dependencies using Poetry:

git clone https://github.com/lahlfors/Reliability-in-Agentic-AI
cd adk-samples/python/agents/financial_advisor

# Install dependencies (includes FinGuard reqs)
cd financial-advisor
poetry install --with dev,deployment

2. Running FinGuard (New)

To run the FinGuard verification suite (Integration Tests):

PYTHONPATH=$(pwd) poetry -C financial-advisor run python3 finguard/main.py

3. Running Legacy Financial Advisor

To run the original agent locally:

PYTHONPATH=. poetry -C financial-advisor run python3 financial-advisor/main.py

References & Further Reading

This project's architecture is informed by the following research in AI safety and governance:

1. AgentGuard: Runtime Verification of AI Agents
   • Koohestani, R. (2025). arXiv:2509.23864
2. The Unified Control Framework: Establishing a Common Foundation for Enterprise AI Governance
   • Eisenberg, I. W., Gamboa, L., & Sherman, E. (2025). arXiv:2503.05937
3. Systematic Hazard Analysis for Frontier AI using STPA
   • Mylius, S. (2025). arXiv:2506.01782

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for more details.