WARDEN – Agent Spend Policy Guard (ASPG)
EXECUTIVE SUMMARY
WARDEN is a policy‑enforcement gateway for x402 autonomous agent payments that implements a deterministic, deny‑by‑default policy engine consisting of 10 sequential gates and a 7‑signal risk scoring system. It ensures that whenever an agent attempts to spend funds, the attempt is cryptographically verified and recorded on‑chain—even when refused—providing irrefutable proof that "The agent tried to spend $2. The Guard refused. The blockchain has no record of the attempt" is mathematically impossible.
PROBLEM STATEMENT
Autonomous AI agents require secure payment intermediaries that prevent unauthorized spending while providing verifiable proof of transaction outcomes. Existing solutions either allow agents direct access to private keys (creating theft risks) or lack cryptographic proof for refused transactions, leaving merchants unable to demonstrate due diligence.
INVESTIGATIVE FINDINGS
- ▸AI agents require constrained spending capabilities to prevent financial harm from malfunctioning or compromised systems
- ▸Blockchain transactions are irreversible, making pre‑transaction validation essential to prevent fund loss
- ▸Financial regulators require proof of both attempted and completed transactions for audit trails
- ▸Systems must provide cryptographic evidence of transaction refusal to protect merchants from false claims
BEHAVIORAL ANALYSIS
Engineered as a zero‑trust intermediary where AI agents never retain access to private keys or signing capabilities. All policy decisions are made agent‑side and cryptographically committed to blockchain before any value transfer can occur, ensuring the agent cannot bypass or spoof guard decisions.
TECHNICAL EXECUTION
Built with Next.js 16 (App Router), React 19, TypeScript, Tailwind CSS v4, PostgreSQL with Drizzle ORM, and the @x402/* v2.22.0 SDK. Implements a defense‑in‑depth policy engine with 10 sequential validation gates (signature, nonce, amount, destination, timing, frequency, volume, policy, security, and destination integrity) combined with a 7‑signal risk assessment model. Integrates with Algorand TestNet and USDC ASA 10458941 for real settlement, using bigint integer arithmetic throughout to eliminate floating‑point precision errors in financial calculations.
SYSTEM ARCHITECTURE
OPERATIONAL FEATURES
- ✓Deterministic deny‑by‑default policy engine with 10 sequential validation gates
- ✓7‑signal risk scoring model (amount, velocity, frequency, geographic, behavioral, temporal, and anomaly signals)
- ✓Real USDC TestNet settlement on Algorand blockchain with on‑chain proof of every decision
- ✓Split‑screen dashboard showing raw agent requests and guard decisions in real time
- ✓WebSocket‑based low‑latency streaming for real‑time updates
- ✓Comprehensive audit trail with cryptographic hashes linking all decisions
- ✓Budget ledger with advisory locks, velocity tracking, and budget windows
- ✓Demo scenarios demonstrating ALLOW, HOLD, and BLOCK outcomes with transaction proofs
TECHNOLOGICAL ASSETS
CURRENT STATUS
Active development with deployed testnet instance. Recent accomplishments include successful completion of the ACTS EDC Brainwave 2026 demonstration, implementation of multi‑sig wallet support for enhanced security, and optimization of the proof‑generation pipeline. Current work focuses on mainnet readiness, additional blockchain integrations, and expanding the policy gate library.
CONCLUSION
Provides cryptographic certainty for autonomous AI agent payments by ensuring every transaction attempt—whether permitted or refused—is immutably recorded on blockchain, eliminating the possibility of undetected spending attempts while maintaining full agent autonomy within defined boundaries.