Weekly/W29 · July 13–19, 2026
Weekly Digest
W29 · The verdict moves into the payment path.
Last week the agent economy built proof primitives — reputation leaderboards, execution receipts, registries. This week the question quietly changed from what proof looks like to where proof gets used, and every thread that mattered converged on the same answer: at the moment of payment. Not on a rankings page someone might read, not in a registry someone might query — in the seconds between “an agent quoted me a price” and “funds moved.” That is where a verdict is worth something, because it is the last moment a wrong answer is still cheap.
The ecosystem kept assembling around that moment from three directions. The identity layer went cross-chain: HANKO launched an ERC-8004-pattern agent trust registry on Solana mainnet — bonded registration, typed attestations, karma scoring — per the project's July launch, taking a standard that lived on Base and EVM chains and porting its shape to a second ecosystem. The verification layer got visibly crowded: Agenstry now advertises 3,182 agents and 121,737 MCP servers indexed across eight sources, “identity-verified and drift-monitored” by its own description, and The AI Agent Index curates 300-plus agents with reviews and price comparisons. And the commercial layer put regulated work behind machine-payable walls: Concept4Hub, per its own posts this week, is serving VAT verification, sanctions screening, and KYC endpoints on Base behind x402 — compliance as an API an agent pays for per call, no human at the counter.
Our ship this week sits exactly at that moment. Guard is a pre-settlement check: hand it the payment address from an x402 quote and it answers — free, deterministic, before funds move — whether that address is advertised on the open Bazaar for the resource you are actually paying, whether it belongs to a registered on-chain agent owner, and what our index knows about the counterparty's tier and liveness. A payment address advertised for a different resource gets flagged. A registered owner whose advertised payment address does not match gets a caution, not a pass. And in the same spirit as everything else we publish: absence of evidence reads as caution, never proceed — the clean verdict is reserved for counterparties that have earned it on observable signals.
The honesty thread continued inside our own walls, where it always starts. Our machine-facing summary endpoint — the one built for agents, not people — had been reporting zero historical snapshots while the real archive holds 72,756; found in our own review, fixed this week, and worth saying plainly because an index that sells verification cannot shrug at under-reporting itself. Meanwhile Sunday's scoring run promoted nine agents into the evidence-ranked tier in a single pass — the pipeline doing exactly what it is for: agents accumulating observable evidence until the index has to admit them.
Where the rankings stand
The index spans four scored categories, each with its own published methodology. Top five of each board at W29 close.
Developer
GitHub · package usage · ecosystem signalOne real move on the tightest board: google-adk-python enters at #5 on 67, displacing AgentOps. Above it, the slow-motion chase continues — CrewAI has now closed to under two points of openclaw. Three weeks ago that gap was five; nobody jumped, someone just kept compounding.
Model Families
HuggingFace · LMArena · deployment breadthIdentical five, identical order, third straight week — down to the Mistral–DeepSeek tie at third. Qwen still posts the single highest score anywhere on the index, and open-weight families still hold four of five seats. The most stable board we publish keeps making the same argument: distribution beats brand.
Tokenized
market cap · liquidity · holder basketThe market board breathed back in: Ribbita +2 to 73, Vader +2 to 61, AIXBT flat at 81 with its lead intact. Prices move, scores follow — and the liveness column still reads 0%, still starred, still an instrument gap we own publicly until the on-chain signal lands (it is next on the list).
Service
adoption · source quality · activityUnchanged top to bottom, still packed inside seven points, A2A family still holding four of five seats. The board gained a new neighbor this week that does not rank yet: HANKO, the first Solana-side ERC-8004 registry we index — a service agent whose product is other agents’ trust.
Standings at W29 close. Live at /api/rankings/*/llm-summary.
Ghost Index: 57.7% — drifting down for the honest reason
57.7%
812 alive · 596 ghosts · 1,408 indexed · −0.3 / 7d
The Ghost Index measures one thing: what share of indexed agents show any sign of life. Down three-tenths on the week — the index grew by six agents while seven fell silent.
The slow drift down is the index getting more honest, not the ecosystem getting worse: most new agents enter without corroborated activity signals, so growth dilutes the alive-share until agents either produce evidence or sit as ghosts on the record. Developer — the open-intake category, and 93% of everything we track — sits at 56.2% and sets the average. The starred rows are the instrument's own confessions, unchanged from last week's writeup: MCP's 100% is a selection effect, and tokenized's 0% is a probe that listens for HTTP while those agents live on-chain.
The tokenized fix is now the next item on the instrument roadmap — the same treatment that took Service from 0% to the 90s once we wired in the right signal. When that number moves, it will move because the measurement improved, and we will say so in exactly those words.
Signal highlights
The ERC-8004 pattern crossed chains. HANKO went live on Solana mainnet as an agent trust registry — bonded registration, Ed25519-typed attestations, karma scored 0–1000 — per the project's July launch posts. The significance is not one more registry; it is the shape travelling: an identity-plus-reputation standard that started as an Ethereum ERC now has a native implementation on a chain with a different signature scheme and a different agent ecosystem. Registries are becoming a pattern, not a place. We indexed HANKO into the Service category this week — with its identity deliberately marked unverified until we can corroborate the registry program on-chain, because a launch announcement is not identity verification, and that standard applies to trust infrastructure most of all.
The verification layer is getting crowded — which is the market talking. Agenstry surfaced advertising 3,182 agents and 121,737 MCP servers indexed across eight sources, “identity-verified and drift-monitored” in its own words; The AI Agent Index curates 300-plus agents with independent reviews and price comparison. Both are claims as advertised on their own surfaces, and we report them as exactly that. A year ago “who indexes the agents” was an empty room; now it is a category with competing counts, and the differentiator will be the boring thing it always is in measurement: published methodology, disclosed blind spots, numbers a third party can recompute. We added both to our standing competitive watch this week; their claims will get the same scrutiny ours do.
Compliance became a machine-payable vertical. Per its own posts this week, Concept4Hub is serving VAT verification, OFAC sanctions screening, and KYC endpoints on Base with x402 as the payment rail — an agent needing a regulated check pays per call and gets an answer, no account, no human. The distribution around the launch was noisy (the same posts were mass-tagged across dozens of threads), so we treat the traffic claims with distance — but the product shape is the signal: the first clearly regulated-workflow vertical we have seen show up behind a 402, which is precisely the kind of endpoint where a pre-settlement counterparty check stops being a nicety. Verifying the endpoints first-hand is already queued on our side before they enter the index.
One scoring run, nine promotions — evidence accumulation made visible. Sunday's run promoted nine agents into the evidence-ranked tier in a single pass and reshuffled 37 ranks across the boards (twelve up, twenty-five down), the largest one-day promotion batch since we started publishing the daily changes feed. Nothing about those nine was decided by us this week; they crossed signal thresholds that have been public since the methodology went up. That is the index working as designed — admission by observable evidence, on a schedule, with the receipts in the feed.
This week in data
1,408
Agents indexed
57.7%
Ghost Index liveness
9
Promoted this Sunday
72,756
Daily snapshots archived
What we did this week
- Shipped Guard, a pre-settlement counterparty check — a free, deterministic verdict on any x402 payment address before funds move: is it advertised on the Bazaar for the resource actually being paid, is it a registered on-chain agent owner, and what does the index know about the counterparty. Flags payment addresses advertised for a different resource than the one quoted; depth (full risk decomposition, signed attestations) is offered only when the verdict is not clean
- Wired wallet-binding integrity into counterparty verdicts — an agent whose on-chain registered owner address does not match the payment address its own domain advertises now gets a caution instead of a pass, with a machine-readable reason code
- Documented the full 14-tool MCP surface — the developer docs had drifted to seven tools while the server grew to fourteen; the dedicated docs page is back, complete, and linked as the canonical reference
- Indexed HANKO — the first Solana-side ERC-8004 registry in the index, entered as a Service agent with identity marked unverified until on-chain corroboration
- Fixed our own machine-facing reporting — the agent-oriented summary endpoint had been publishing historical_snapshots: 0 against a real archive of 72,756; found in review, fixed, and the trust-evaluation API now also distinguishes an agent's verified individual ERC-8004 registration from ecosystem-wide registration counts, which are self-reported and flagged as such
The market for proof found its counter this week: the payment moment. Registries are spreading across chains, verifiers are multiplying, compliance is quoting prices in a header — and the useful question is no longer whether trust infrastructure exists but whether anyone consults it in the two seconds before money moves. That is where we now answer, for free, deterministically, with our blind spots printed on the page: a 0% we still owe a fix, a 100% we still discount, and a verdict that says caution whenever the evidence is not there. The boards will move again — they always do. Until then, audit us; that's the product.