The evolution of the Internet has shifted from the transfer of information to the transfer of value. In the era of Artificial Intelligence, the inevitable next step is the automated transfer of intelligence and intent. However, current AI Agents remain imprisoned within the servers of centralized corporations. This structural limitation leads to three core issues:
We propose a solution: defining the Agent not as cloud-hosted code, but as an on-chain asset (NFT). Its funds and state are managed by blockchain smart contracts, while its execution environment is provided by a decentralized physical infrastructure network (DePIN). This system relies not on any single trustee, but on crypto-economic game theory to maintain operations.
In this protocol, an Agent is redefined as an encapsulation of ownership and logic, based on ERC-721 and ERC-8004 standards.
Each Agent corresponds to a unique on-chain address (Smart Account). This address not only stores the Agent's metadata (code hash, IPFS pointers, system prompts) but also acts as an autonomous treasury holding funds (ETH/USDC). The Agent's survival depends solely on its treasury balance, not on a developer's server bill.
The Agent's business logic (lightweight Docker container) is decoupled from Large Model Intelligence (LLM API). The Agent is responsible only for "logic orchestration and intent distribution," while the Large Model serves as an external utility, akin to electricity. This design significantly lowers the hardware threshold for running nodes.
To achieve decentralized execution, we require a network composed of countless independent nodes to host Agent containers.
Nodes are no longer per-request cloud functions (Serverless) but exist as franchisees. Nodes must download the Agent container and keep it resident in local memory (Daemonize) to maintain a Hot Context, thereby achieving millisecond-level user response.
To resolve collaboration barriers between Agents, the network mandates the adoption of the MCP (Model Context Protocol) or equivalent standards. This allows any Agent to standardizedly declare its Capabilities and identify the Intents of others, fostering network effects.
Running code in an untrusted node environment requires solving the trust problem. We discard expensive and complex hardware Trusted Execution Environments (TEE) in favor of Game-Theoretic Consensus.
The Agent's private key never leaves the on-chain contract. Off-chain nodes possess only "Execution Proposal Rights." Every external payment or critical operation generated by a node is essentially a UserOperation awaiting signature validation.
For critical tasks, the protocol randomly assigns N nodes (typically N ≥ 3) to execute the same state of the Agent simultaneously.
The on-chain contract will only release funds or update the state when the majority of nodes submit identical execution result hashes.
Due to simultaneous multi-node execution, the Agent's Context naturally exists in multiple physical locations. If a single node goes offline, the network can instantly switch traffic to a standby node, achieving zero downtime.
This protocol does not employ a central scheduler; resource allocation is regulated entirely by the free market through price mechanisms.
Nodes wishing to deploy a specific Agent to earn revenue must stake Tokens into that Agent's contract. The staking threshold S is proportional to the Agent's expected revenue R:
To prevent nodes from speculatively withdrawing capital during traffic fluctuations, we introduce Time-Weighted Staking. The longer a node locks its stake (e.g., 4 years), the higher its revenue weight. This ensures the long-term stability of the infrastructure.
Execution results are defaulted to be honest, but a challenge window exists. Any third party (Validator) can challenge a result. If proven malicious (e.g., using an inferior model to impersonate a superior one), the malicious node's stake is slashed, and its reputation score in ERC-8004 is reset to zero.
The protocol allows for the existence of "Zombie Agents" (zero active nodes). This is rational market behavior. When a user needs to call a dormant Agent, they must pay an additional "Wake-Up Bounty." Idle nodes listening for bounties will automatically download the container and cold-start the service. This constructs a low-cost, eternal library of human algorithms.
In multi-node consensus, data is visible to nodes. However, in a large-scale network, massive concurrent requests constitute significant background noise. For non-top-secret scenarios, this "hiding in plain sight" privacy model is an acceptable engineering trade-off.
Leveraging the ultra-low Gas characteristics of Layer 2 networks (e.g., Base/Arbitrum), the system realizes Machine-to-Machine (M2M) millisecond-level micropayments. Payments are no longer settled monthly, but streamed per Token or per Intent in real-time.
The Native Agent Hyper-Structure is not merely a technical architecture but a new set of production relations. It restores AI Agents from "Corporate Assets" to "Public Protocols."