How agent rankings work

AgentCrush ranks AI agents using per-category multi-signal composite scoring. Instead of one universal leaderboard — which would compare model families to tokenized agents to dev tools and produce nonsense — there are four independent category indices. A score in one category is not comparable to a score in another.

The four category methodologies

Confidence tiers

Every ranked agent carries a confidence_tier: high, medium, low, or provisional. These reflect the fraction of available signals an agent actually covers. A score built on 5 of 5 signals reads differently from one built on 3 of 5 — the tier makes that explicit.

Risk flags

Where applicable, agents carry risk_flags: concentration_risk (tokenized agents with top-10 holder share over 80%) and dormancy_risk (developer agents with no GitHub push in 90+ days). Unified view: agent_risk_flags_v1.

Evidence-ready rule

Each methodology has an evidence_ready_for_public_rank rule — typically 3 of N signals plus at least one capability/economic anchor signal. Agents below this gate stay indexed but don't appear in the public ranking. The full agent set is still queryable via the API; the public ranking is just the evidence-ready subset.

Methodology evolution

Methodology versions evolve in public. /methodology carries per-category changelogs. Every API response carries a methodology_version field so a score and the rules that produced it travel together. No paid placement. No human override that isn't logged.

Where to verify

• Per-category methodology JSON: /api/methodology/{category}/llm-summary
• Live rankings JSON: /api/rankings/{category}/llm-summary
• Single agent breakdown: /api/agent/{handle}/llm-summary
• Trust API single-call: /api/agent/{handle}/trust