HadAgent: Harness-Aware Decentralized Agentic AI Serving with Proof-of-Inference Blockchain Consensus
New consensus mechanism uses LLM inference tasks to secure a decentralized network, turning wasted compute into productive AI agents.
A team of researchers including Boyang Li has published a paper on HadAgent, a novel framework designed to solve two major tech problems at once: the energy waste of cryptocurrency mining and the soaring demand for GPU compute to run large language model (LLM) agents. The core innovation is Proof-of-Inference (PoI), a blockchain consensus mechanism that replaces the arbitrary cryptographic puzzles of Proof-of-Work (PoW) with useful, deterministic LLM inference tasks. Nodes on the network compete to create new blocks by performing these AI computations, effectively turning the massive computational effort required for blockchain security into a productive engine for decentralized AI agent serving.
HadAgent's architecture is built for security and efficiency in this new paradigm. It organizes transaction data into a three-lane block body with separate DATA, MODEL, and PROOF channels, each secured by an independent Merkle root for granular integrity checks. A sophisticated two-tier node system classifies participants as trusted or non-trusted based on historical performance, allowing trusted nodes to serve inference results optimistically for speed, while others undergo full verification. A continuous "harness" layer monitors all nodes through heartbeat probes and anomaly detection, creating a self-correcting system that can demote malicious actors and promote reliable ones.
According to the prototype's experimental results, the system demonstrates remarkable performance. It achieved a 100% detection rate for tampered records with zero false positives, and validation operations for records and network hubs occurred in under a millisecond. The trust management system proved effective, excluding adversarial nodes within just two consensus rounds and promoting honest participants to trusted status within five rounds. This suggests a robust and responsive framework for maintaining network integrity while performing useful AI work.
- Introduces Proof-of-Inference (PoI), repurposing blockchain compute from crypto mining to running deterministic LLM tasks for AI agents.
- Uses a three-lane block structure (DATA, MODEL, PROOF) and a two-tier trust system to optimize security and serving speed.
- Prototype demonstrated 100% tamper detection, sub-millisecond validation latency, and effective isolation of bad nodes within two rounds.
Why It Matters
It could create a sustainable, decentralized marketplace for AI agent compute by turning security overhead into a valuable service.