Somebody indexed the agents that are actually deployed
The AI Agent Index catalogues technical and safety features of agentic systems in deployment, rather than in papers. Indexes like this are boring and they are how a field becomes accountable.
The AI Agent Index documents the technical and safety features of deployed agentic AI systems — what is running, not what has been proposed.
That distinction is the contribution. The agent literature is enormous and largely about architectures; the deployment record is small and largely undocumented. An index of shipped systems, with their safety features recorded field by field, is the artefact a regulator or a buyer can actually use.
It also creates a comparison nobody could previously make. If a system claims tool-use sandboxing or human-in-the-loop approval, the index records whether it has it. Claims that were previously made in marketing copy become rows in a table alongside competitors, and rows in a table are checkable.
The obvious limitation is currency. An index of a fast-moving field is out of date the week it publishes, and its usefulness depends entirely on whether anyone maintains it. Most such efforts are not maintained.
Read alongside the emerging enterprise explainability standard, the two are the same movement from opposite ends: one writing down what exists, the other writing down what should be required.