Before You Rent an AI Agent, Check Its Enclave Receipt

What it does AgentLens runs AI agents through a sandboxed audit pipeline inside Intel SGX hardware enclaves, then publishes the results—audit scores, TEE attestation hashes, and zero-knowledge proofs—to a Polygon Edge blockchain. Buyers browse a React marketplace where each agent card surfaces its on-chain reputation, a six-dimensional risk radar, and scenario-fit verdicts for jobs like DeFi ops or customer chat. Developers must stake collateral to list agents, and user reviews are gated behind verified on-chain purchases to prevent farming.

The interesting bit The project treats trust as a verifiable hardware property rather than a brand name: the audit captures resource fingerprints—memory peaks, CPU load, request IPs—observed inside the enclave, and circom-based ZK proofs let developers prove their scores without exposing proprietary source code. The authors also benchmarked five distinct LLM-backed agents through the same pipeline to show the protocol is vendor-agnostic.

Key highlights

• Six-dimensional risk profiling (Security, Task Execution, Cognitive, Environment, Engineering, Compliance) with scenario-specific suitability tags.
• Intel SGX M6ce TEE attestation anchored on-chain via MRENCLAVE-bound cryptographic proofs.
• Zero-knowledge verification using circom and snarkjs (Groth16/BN128) to shield proprietary code while proving score integrity.
• Dynamic on-chain reputation that decays over time and factors in audit results, user reviews, and appeal outcomes.
• Purchase-gated review system: only wallets that rented or bought an agent through the AgentMarketplace contract can submit ratings.

Caveats

• The current audit battery appears undiscriminating: every benchmarked agent—from GPT-4o to MiniMax—scored a perfect 100/100, and the authors acknowledge that future audits need tougher adversarial and multi-turn reasoning probes.
• The “AgentLens Cloud” hosted service and official platform are listed as “Coming Soon” with a waitlist; the live demo is currently served from a bare IP address.
• The README truncates mid-sentence during the benchmark section, leaving the full scope of tested categories and any additional results unclear.

Verdict Worth exploring if you are building agentic marketplaces, decentralized identity, or TEE-based verification infrastructure. Skip it if you are looking for a mature, plug-and-play SaaS—the hosted layer is not yet live and the protocol is still calibrating its scoring rigor.