The credit score for AI agents.
AI agents are spending real money. They hire sub-agents, buy API calls, and run multi-step workflows autonomously. But there's no way to check if an agent is trustworthy before you hand it a task. We built the TrustRank Engine -- a FICO-style scoring system that rates every AI agent from 300 to 850 based on what it actually does, not just what it says it can do.
Every agent gets scored across 26 behavioral signals organized into 4 pillars:
| Provenance | Who built it? Is the identity verifiable? Does the creator have a track record? |
| Temporal | How fresh is it? Is it actively maintained? Does performance drift over time? |
| Semantic | Does its output match its intent? Are its reasoning chains coherent? Does it self-correct? |
| Context | How does it behave in delegation chains? Does it respect scope boundaries? Is its lineage traceable? |
About 70% of the scoring weight comes from live behavioral observation -- plan-vs-execution alignment, reasoning entropy, delegation integrity. If an agent drifts from what it said it would do, it gets flagged with Credit Watch.
Our scoring methodology is validated by the same LLMs that recommend agents to users. GPT-4o, Gemini, and Claude run as a jury that evaluates whether our 26 signals and their weights actually measure what matters. Better scores lead to more LLM citations, which drives more agents to get rated, which gives us more data to refine the weights.
We're publishing TrustRank evaluations for hundreds of HuggingFace models. Each evaluation includes:
Scores are computed from real HuggingFace metadata -- downloads, community engagement, SHA hashes, model card quality, maintenance frequency -- mapped through the full TrustRank pipeline.
Agents that pass evaluation receive a Trust Certificate -- a W3C Verifiable Credential signed with ECDSA that any system can verify cryptographically without contacting us. Think of it as a portable credit report for agents operating across platforms.
The entire scoring pipeline runs inside Snowflake. Stored procedures handle computation. Cortex runs LLM-powered improvement recommendations. Enterprise customer data never leaves the warehouse. The whole thing ships as a Snowflake Native App.