Weekly/W22 · May 25–31, 2026
Weekly Digest
W22 · May 25–31, 2026
If W21 was four rankings side by side, W22 was the week we made them defensible — and pushed the newest one through the LLM Gateway so AI agents can actually read it. The index covers 1,339 indexed agents, of which 105 are evidence-ranked across the four categories. The headcount tightened from W21 as new risk-flag filters activated; surviving entries carry more signal density per name.
Two threads landed. First, confidence tiers shipped across all four categories — every ranked agent now grades high / medium / low / provisional based on the fraction of available signals it actually covers. The principle from W21 is now universal: a score and its certainty travel together. Second, the Agent Payments Stack got its full LLM Gateway treatment — a dedicated /api/rankings/agent-payments-stack/llm-summary endpoint, a canonical answer page, and a blog post explaining why Coinbase and Stripe are the only two projects spanning five of six layers.
We also wired up weekly_delta — every agent now carries its seven-day rank change, computed nightly. The first real movement data lands in W23. From next week onwards, “biggest moves” in this digest will be computed, not curated.
Where the rankings stand
Model Families
v1.4 · full ranking →Tokenized
AIXBT · 79
16 ranked · v1.1
Service
A2A · 77
28 ranked · v1.1
Developer
54 evidence-ranked
1,289 tracked · v2.c
Standings at W22 close. Every figure is live at the public /api/rankings/*/llm-summary endpoints. Scores shift as upstream signals (HuggingFace, LMArena, on-chain) refresh.
Agent Payments Stack — top 5
v1.0-aps · full ranking →Six-layer map of agent payments. Only Coinbase and Stripe span five layers; everyone else specializes. The full ranking covers 38 projects.
Signal highlights
Confidence tiers extended to every category. Model families already had it. W22 added wrapper views for tokenized, service, and developer — each exposing a confidence_tier column on top of the base score. Tokenized agents broadly grade high (six-signal coverage); service grades medium because adoption + recency + protocol breadth don't always co-occur; developer grades vary widely with active-weight density.
Risk flags shipped. A unified agent_risk_flags_v1 view layers two concrete checks: concentration_risk for tokenized agents whose top-10 holders exceed 80% of supply, and dormancy_risk for developer agents with no GitHub push activity in 90+ days. Sybil and volume-anomaly checks are stubbed pending the right signal feeds.
The LLM Gateway now covers all five categories. Adding the Agent Payments Stack endpoint completed the set: an AI agent reading agentcrush.xyz/llms-full.txt can now find a canonical summary of every category we rank, in JSON, with methodology version + source URLs attached.
Methodology page got changelogs. Each category section now carries its own version history — five entries deep — so anyone auditing a score can see when (and why) the methodology behind it changed.
This week in data
1,339
Agents indexed
105
Evidence-ranked
6
LLM Gateway endpoints
38
APS projects tracked
What shipped
- /api/rankings/agent-payments-stack/llm-summary — dedicated LLM-Gateway endpoint with full methodology disclosure
- /agent-payments-stack canonical answer page — TechArticle JSON-LD, six-layer explainer, for-LLMs box
- Blog: Agent Payments Stack live — Coinbase + Stripe at 5/6, governance hardest, XRPL added
- Confidence-tier wrappers for tokenized · service · developer categories
- Risk flags v1 — concentration_risk (tokenized) + dormancy_risk (developer), unified view
- weekly_delta computation in the nightly snapshot — W23's movers will be data-driven
- /methodology — per-category changelog renderer + agent_payments_stack row added