Loops HouseNEAR Agent Census

NEAR HACKATHON
Links
Team
1 member- DAOwner
Daodu Destiny
Overview
NEAR's pitch is that software, including AI agents, will transact on it, but nobody publishes a measured number. The NEAR Agent Census is a re-runnable investment-thesis tool that estimates how much of NEAR's activity is software. It draws a seeded random sample of 2,370 blocks (2022-01 to 2026-11, 6,240 transactions), resolves each transaction's real initiator (relayed NEP-366 meta-transactions are attributed to the user, with the relayer tracked as fee payer), profiles that account from its own on-chain history and scores it with an explainable rule list. Result: over 2026-06 to 2026-11, 61% of sampled transactions were initiated by software-like accounts (95% CI 57%-65%; 57%-63% across classifier cut-offs), which initiated 74% of fees and paid 83%. But in absolute terms total activity fell from 64 to 8 tx/s and automated activity from 15 to 5 tx/s, so the share rose because other activity fell faster. On a held-out test split the classifier finds 85% of labelled automated accounts and keeps 99% of labelled humans human (labels are biased toward self-declared bots and engaged humans, so that is an optimistic ceiling). Each claim of the thesis gets a verdict decided by code from confidence intervals (1 inconclusive, 3 descriptive, 2 not supported, 7 supported), including ones that cut against it. The dashboard lets anyone verify: random sampled transactions link to an explorer, and any account's verdict shows the exact rules that fired. NEAR fits because per-account history is public, meta-transactions split who acts from who pays, function-call keys shape agent behaviour, and NEAR Intents traffic is measured. One command, python -m census reproduce, rebuilds every number offline. Automated includes AI agents but cannot isolate them. Not investment advice.