AgentPass: Autonomous AI Finance Governance and Intent Firewall

Inspiration

As autonomous AI agents (built on frameworks such as CrewAI and powered by models such as Gemini 2.0 Flash) evolve from passive assistants to active economic actors capable of making purchases and executing workflows, a massive security gap emerges. Without strict financial guardrails, LLM agents are vulnerable to prompt injection, hallucinated checkouts and runaway spending. AgentPass was built to solve this, as a microsecond policy enforcement and semantic intent firewall between autonomous agents and financial rails like PayPal Sandbox.

What It Does

AgentPass is the full admin control plane + execution gateway for AI commerce autonomy:Agent Registry & Scoped Credentials: API keys scoped agp_live_ with capabilities ecommerce:search and ecommerce:purchase.
Deterministic Spend Controls: Enforces hard limits including Max Per-Transaction, Daily Spend Ceilings, Monthly Budgets and Human Approval Thresholds.
Semantic Intent Verification: Converts natural language user instructions into positive priorities and negative constraints, so that conflicting purchases are avoided (e.g. blocking gaming laptops when college portability is needed).
Multi-Agent Delegations: . Supports parent-child authorization graphs (detects cycles, budget limits on subsets).
Emergency Circuit Breakers: Features a Kill Switch to immediately remove financial permissions from individual agents or the entire platform.
Tamper-Evident Audit Logs: Maintains a cryptographically chained SHA-256 audit trail of all policy evaluations, approvals and settlements

How We Built It

Frontend: Built with React, Tailwind CSS and Vite. Responsive dashboard layout with real-time risk scoring, velocity graphs and enforcement breakdown charts. Backend & Security: FastAPI, PostgreSQL, Redis, Docker, Argon2id credential rotation, PayPal Sandbox webhook receiver rails.
AI & Governance Engine: Integrates CrewAI and Gemini 2.0 Flash for real-time assessment of intent consistency and multi-signal risk scoring.

Challenges we faced Connecting natural language and deterministic limits:

Translating loose human instructions (e.g. “Prioritize battery life, under $1000”) into tight runtime constraints without breaking the agent’s autonomous workflow required careful prompt structuring and robust fallback parsing.
Low-Latency Policy Enforcement: Spend checks, intent verification, and risk scoring must operate within microsecond windows prior to the order payloads being sent to the PayPal Sandbox REST API gateway.

Technical Architecture & Risk Formula

AgentPass computes transaction authorization probability and risk scoring using multi-signal heuristics. For instance, risk scoring velocity and transaction limits are evaluated using models represented as:

$$RiskScore = \alpha(\text{Amount}) + \beta(\text{VelocityCount}) + \gamma(\text{IntentConflict Penalty})$$

Where thresholds are deterministically evaluated at the gateway prior to initiating PayPal capture settlements.

Built With

Share this project:

Updates

Submission history