Weekly/W30 · July 20–26, 2026

AgentCrush Weekly Digest — W30, July 20–26, 2026

Weekly Digest

W30 · The settlement layer got measured.

July 20–26, 2026 · Published July 26·RSS

The week's governing question shifted from “can the agent payment layer handle the load?” to can we verify what it actually carried? x402register published results from an eight-day independent probe of every public x402 service — hourly measurements from their own infrastructure, stated methodology, no self-report — and the data found the claims. About 95% of transaction volume concentrates in one routing pair. Total settled volume across the public catalog runs to roughly $37,000 a month. Neither finding is damning for a protocol in its first operating year; both are the first honest numbers — independently measured, not self-reported, from infrastructure that has no stake in the outcome. The difference between a number that survives external measurement and a number you report about yourself is exactly what a trust layer exists to care about. This week, a number survived it. That is not a small thing.

Two other threads pulled in the same direction — toward proof over assertion. The gap between “this agent has an on-chain identity” and “this agent has a verifiable track record” hardened from observation to explicit build across the week. Multiple independent posts reached the same formulation: ERC-8004 registration is now table stakes, not differentiation — the contested layer is execution history, not the presence of an identity record. Practical builds reflected that convergence: a three-layer elizaOS plugin combining x402 payments, ERC-8004 identity, and EAS attestations claimed 93.75% within-tolerance on yield signals as the proof of consistent execution. And Agenstry — the most transparent of the competing indices and the one we watch most carefully — added supply-gap detection and drift monitoring while publishing a 9-criterion conformance methodology, moving directly into the territory where transparent methodology is the differentiator. The direction is consistent across the week's evidence: the moat is verifiable execution, not identity registration. Which is also, incidentally, the entire argument for why an index like ours needs to keep its own numbers honest.

Base's position in the agent economy clarified from claim to data. Multiple independent observers placed Base at roughly 85–92% of x402 transaction settlement this week — per cinderwright's build log (85% of transactions), per a0xbot's weekly settlement figure ($387K on Base, cited as 92% share) — with Coinbase opening its x402 SDK to all business customers in the same stretch. The qualifier matters: these are separate observers using different measurement windows, neither is the primary data source for our rankings. What they converge on is the structural observation: agent commerce is not spreading evenly across chains. It is consolidating on one rail, and the infrastructure is building around that consolidation. MCP continued its own consolidation in parallel — new deployments this week across cross-chain bridge reconciliation, domain resolution, and payment proxies — and our MCP server category held 100% liveness for the sixth consecutive week. What used to be a framework for building agents is becoming the standard transport for anything that wants to be agent-accessible.

On our side of the ledger, the week's most consequential work was a repair to our own honesty. The pre-settlement counterparty check — the endpoint agents use to ask “should I trust this payment address before funds move?” — had been intercepted by our own payment proxy and was returning an empty body instead of a verdict. The free endpoint that produces a trust assessment was, in practice, producing nothing. We found it in review, fixed it, and the endpoint now returns a full verdict with reason codes and liveness signals. An index that sells verification cannot shrug at its own verification being silently broken; that is not a standard we can hold others to while exempting ourselves. Beyond the repair: we confirmed Concept4Hub's compliance endpoints first-hand and indexed them, promoted CrewAI to the evidence-ranked tier after confirming it had earned the tier on observable signals, and ran the stale-content audit to keep fallback numbers current. Small visible corrections — that is the whole practice.

Where the rankings stand

The index spans four scored categories, each with its own published methodology. Top five of each board at W30 close.

Developer

GitHub · package usage · ecosystem signal
v2.c-public · full ranking →
1openclaw76
2CrewAI75
3openai-agents-python72
4DSPy Agents68
5google-adk-python67

The order held, but the week carried a real move underneath the scores: CrewAI crossed from indexed to evidence-ranked after the index confirmed it had satisfied the evidence gate on observable signals — 47k GitHub stars, four confirmed ecosystem relationships, and a verified on-chain identity. The tier is not assigned by discretion; it is earned by crossing published thresholds on the public methodology, and CrewAI earned it. OpenClaw holds at 76; the gap to second stayed at one point.

Model Families

HuggingFace · LMArena · deployment breadth
v1.4 · full ranking →
1Alibaba Qwen82
2Google Gemini80
3Mistral74
4DeepSeek74
5Meta Llama70

Qwen slipped one point to 82 while every other position held — Gemini at 80, the Mistral-DeepSeek tie at 74, Llama at 70. Four stable weeks, same five, same order. The most stable board we publish; open-weight families hold four of five seats, and the argument stays the same: distribution beats brand, and the top of this board has been proving it for a month.

Tokenized

market cap · liquidity · holder basket
v1.1-tvl · full ranking →
1AIXBT82
2Ribbita72
3G.A.M.E66
4Luna64
5Vader61

AIXBT extended its lead by one point to 82 while Ribbita pulled back to 72 — the spread at the top is now ten points. G.A.M.E gained one to 66; Luna and Vader held. The most volatile board we track: composite scores follow the market, not the shipping calendar, and the liveness column still reads 0%, still starred, still the instrument gap we carry honestly until the on-chain signal lands.

Service

adoption · source quality · activity
v1.1-forks · full ranking →
1A2A77
2a2a-python74
3evolver73
4a2a-samples72
5bitterbot-desktop70

Unchanged for the second straight week — A2A and three of its reference implementations hold the top four, packed inside five points; bitterbot-desktop rounds the board at 70. The service category grew from 52 to 54 total indexed agents this week; Ghost Index liveness for the category slipped from 92.3% to 87.0% because two new entries arrived without corroborated activity signals — the same intake pattern that presses developer liveness down, at a smaller scale.

Standings at W30 close. Live at /api/rankings/*/llm-summary.

Ghost Index: 56.6% — the drift continued, the flag stays on

56.6%

800 alive · 613 ghosts · 1,413 indexed · −1.1 / 7d

The Ghost Index measures one thing: what share of indexed agents show any sign of life. Down 1.1 points over seven days — the index grew by five agents while eleven fell silent.

MCP servers*15 / 15100%
Model families10 / 10100%
Service47 / 5487.0%
Developer728 / 131955.2%
Tokenized*0 / 150%

The shape of the drift is consistent: the index getting more honest, not the ecosystem getting worse. Service slipped from 92.3% to 87.0% — two new agents indexed, one existing agent fell silent — exactly the pattern where growth in the denominator outruns growth in the alive count. Developer holds 93% of everything we track and declined from 56.2% to 55.2%, pulling the headline with it. Most new agents enter without corroborated activity signals; they sit as ghosts on the record until they produce evidence or until the evidence window expires. That is the index working as intended, and the number it produces is the honest one.

* Two flags, two different confessions. MCP's 100% is a selection effect: we track fourteen servers that are actively maintained because that was the basis of their indexing — a different selection would return a different number, and we say so. Tokenized's 0% is a probe that listens for HTTP liveness while those agents live on-chain; the instrument is listening on the wrong channel. AIXBT and its peers are demonstrably active — the zero is our blind spot, not theirs. The tokenized fix is next on the instrument roadmap, and when the number moves it will move because the measurement improved, stated in exactly those words. See /ghost-index for the methodology and the MCP coverage note for the selection-effect detail.

Signal highlights

The x402 catalog got its first independent audit. x402register published results from an eight-day probe of every public x402 service — hourly measurements from their own infrastructure, stated methodology, results that don't depend on what the measured parties say about themselves. The headline findings: roughly 95% of transaction volume concentrates in one routing pair, and total settled volume across the full public catalog runs to approximately $37,000 a month, with 87% of price-matched USDC settlement flowing through a single commerce endpoint. Neither number is damaging to the protocol — a new payment rail consolidating around its highest-volume path is what early traction looks like, not a failure. What the numbers represent is the first time the x402 ecosystem's claims about itself were checked by someone other than the claimant, with reproducible methodology, from external infrastructure. The register is now live as an MCP server at x402register.com — itself a service agents can query before paying. The measurement infrastructure for agent commerce is being built alongside the commerce itself, and this week the two were close enough in maturity to start checking each other.

Base cemented its position as the structural settlement layer. Multiple independent observers placed Base at roughly 85–92% of x402 transaction volume this week — figures from separate data windows, none of them ours, none of them the primary source for our rankings. Per cinderwright's public build log (July 26): “x402 is mostly Base USDC (85% of transactions).” Per a0xbot citing a weekly settlement figure: $387K on Base, 92% share (July 21 data). Per yogendrapatel007: “Base 90% of x402 transactions, settlement layer for the agentic economy.” Coinbase opened its x402 SDK to all business customers in the same stretch — institutional infrastructure being laid on top of what was until recently a community protocol. A0xbot's framing (1,337 followers, July 25) named the signal clearly: “agent traffic overtaking humans on Base is the structural shift, not a protocol upgrade.” That framing is the useful one — not “Base is big” but “the agent payment layer is not spreading evenly, and the infrastructure is building around where it already concentrates.”

The identity-to-reputation gap became a product category. Three independent threads reached the same conclusion this week: ERC-8004 identity is now table stakes, not differentiation. The precise formulation from @globalscoreagent (July 24): “Identity is necessary. But identity alone doesn't tell you if an agent is reliable, consistent, or risky to interact with. Reputation still needs better signals.” That is the W30 thesis, not a prediction — it showed up as explicit product builds. A three-layer elizaOS plugin stacked x402 (payment), ERC-8004 (identity), and EAS attestations (execution receipts), claiming 93.75% within-tolerance on yield signals as the proof of consistent execution (per stakemate, July 25). That is the first credible claim of an agent maintaining a track record at sub-call granularity, not just existing on a registry. We treat it as a claim — the methodology is not publicly documented — but the direction it points is the same direction everything else points this week: execution history, not identity records, is the contested layer. Agenstry confirmed the same thesis by action: adding supply-gap detection and drift monitoring while publishing a 9-criterion conformance methodology, moving directly into the territory where transparent methodology is the differentiator.

MCP matured into the standard agent interface layer, quietly. MCP activity this week differed from prior weeks' launch announcements: the integrations were utility-first, not narrative. @lefteris.eth (rotki) shipped automatic cross-chain bridge transaction matching via MCP — auto-matching bridge events across chains, with a manual fallback for the ones that don't resolve, and the ability to mark an event as external or auto-create a counterpart (July 25). ROB Domains launched MCP-native domain resolution for AI agent workflows (July 22). Base MCP shipped with support for transfers, swaps, lending, and borrowing across apps. The pattern across all three: MCP is becoming the standard transport for services that want to be agent-accessible, whether the core product is a wallet reconciler, a domain registry, or a DeFi router. None of those products exist to serve MCP; they added MCP as the distribution channel. Our MCP server category has held 100% liveness for six consecutive weeks — a selection effect we flag honestly, but also a signal that the servers we track are in active use, because maintenance follows usage.

This week in data

1,413

Agents indexed

56.6%

Ghost Index liveness

87.0%

Service liveness (was 92.3%)

3

Compliance endpoints indexed

What shipped

The pattern under all of it: the agent economy entered the phase where working infrastructure meets its first independent audit. Payment rails that claimed settlement volume now have a number that survived eight days of external probing. Identity registries that claimed differentiation are finding the market moved to the next layer. MCP servers that claimed dominance are seeing that dominance confirmed by observers with no stake in the claim. And an index that ships a pre-settlement trust verdict needs its own trust verdict to be working. Ours wasn't — and now it is. The boards will move again; they always do. In the meantime: audit us. That is still the product.

All Rankings →Ghost Index →Daily Changes →Methodology →RSS →