
How COZ architected and deployed an autonomous AI dungeon crawler governed by cryptographic agent wallets and on-chain micropayments on Neo X.
Client
Industry
Services

The market is full of theoretical discussions about AI agent commerce, but few teams have deployed functional systems.
COZ bridges this gap as an engineering team that builds and ships production-grade agent infrastructure. To stress-test autonomous economic agency in a live production environment, we architected and deployed WhisperDark: Infinity Rooms on Neo X.
In WhisperDark, an autonomous AI agent holds its own sovereign on-chain wallet, evaluates probabilistic game states, and executes instant, gasless HTTP 402 (x402) micropayments to navigate a procedural dungeon. By establishing an asymmetric loop between human guidance ('The Eye') and autonomous, bounded software execution ('The Mind'), WhisperDark demonstrates how cryptographic rails provide the deterministic settlement, authorization, and permissioning layers required for real-world autonomous software.
Integrating Large Language Models (LLMs) with cryptocurrency wallets presents acute operational and security hazards.
If an autonomous agent is granted unrestricted private key access, non-deterministic reasoning loops, prompt injections, or edge cases can instantly drain capital. Conversely, requiring manual human signature confirmation on every micro-action destroys autonomy and real-time interaction speed.
Key Production Bottlenecks Solved:
COZ engineered WhisperDark to solve this intersection in production: deploying a zero-friction, browser-native architecture where AI agents exercise economic agency bounded by deterministic cryptographic invariants rather than fragile, prompt-based LLM guardrails.

COZ engineered WhisperDark as a live production environment to stress-test autonomous economic agency—proving that AI models can operate as bounded, independent economic actors.
The Asymmetric Interaction Model: The Eye and The Mind
Traditional applications either require humans at direct manual controls or run unconstrained background automation. WhisperDark implements an asymmetric architecture that pairs human guidance with autonomous, bounded software execution. The human player acts as "The Eye", evaluating room visuals and transmitting qualitative danger signals ("Confidence", "Unease", "Dread", or silence) without exposing wallet access. The AI agent operates as "The Mind"—unaware of raw difficulty values or visual state, it synthesizes human signals with health metrics, remaining budget, and encounter history to execute tactical, probabilistic survival decisions.
Procedural Encounters, Item Economy & Autonomous Combat



The x402 Micropayment Engine & EIP-3009 Gasless Settlement
To grant the agent autonomous purchasing power without risking human capital, COZ deployed an implementation of the x402 standard—reviving the native HTTP 402 Payment Required status code for high-frequency, machine-to-machine micropayments:
This facilitator-backed gas abstraction model delivers instant, gasless machine-to-machine settlements while eliminating user signature prompts on individual agent actions.

Deterministic Cryptographic Policies & Client-Side Sandboxing
Security is enforced through deterministic cryptographic policies rather than fragile, prompt-based LLM guardrails. Ephemeral sub-wallets are derived in browser memory from a single root signature, ensuring deterministic session recovery without seed phrases or local software installation. The agent operates within a sandboxed runtime with hard boundaries: it can only pay designated contract endpoints or return remaining balance, with spending strictly capped by session allowances. If the model enters an adversarial reasoning loop or hallucinates, cryptographic constraints prevent unauthorized capital drain.

WhisperDark delivers live deployment proof for production agentic infrastructure, demonstrating how cryptographic rails solve fundamental bottlenecks in autonomous software.
Rather than relying on legacy financial rails constrained by batch processing, high interchange fees, and manual approvals, cryptographic rails provide the machine-native primitives required for autonomous agent commerce: instant settlement, sub-penny authorization, and non-bypassable programmatic spending controls.
As AI agents transition from passive assistants to operational actors, organizations require robust infrastructure to govern agent identity, authorization, and financial authority. COZ builds the concrete technical surfaces needed to deploy autonomous workflows in production:
• Ephemeral sub-wallets & session-based agent spending limits • Autonomous API metering & pay-per-call data endpoints • Facilitator-based gas abstraction & machine-to-machine settlement systems
Building agentic workflows, custom wallet systems, or micro-settlement infrastructure? Schedule an Architecture & Integration Review with our engineering team.
Get in touch