Weekly/W26 · June 22–28, 2026
Weekly Digest
W26 · We made our own numbers checkable
W26 was the week roughly 40,000 agents went to work. Virtuals, Ondo Finance and Treasures put more than 430 tokenized stocks within reach of a swarm of on-chain agents, and “agents as traders” went from thesis to live data almost overnight. The loudest post of the week put it bluntly: “what happens when 40,000 AI bots start trading stocks with no human in the loop?” But the more interesting reaction wasn't awe — it was a question. Of those forty thousand, which are economically real, and which can you trust to move money? That question — not whether agents can pay, but whether you can verify what's paying you — was the whole week.
You could watch it consolidate in real time. Across Farcaster and X, the x402 conversation finished its shift from “can agents pay?” to “can I trust what's paying me?” Three sub-threads hardened into one layer: schema-validated, deterministic contracts; custody and signer architecture; and endpoint risk-scoring. ERC-8004 identity went multi-chain in the same stretch — BaseHub registration is live, and Vyper's Arc testnet wired ERC-8004 identity, x402 settlement and Circle payment flows into a single deployment. Identity is now table stakes. The contested layer is proof — liveness, reputation, receipts.
So we spent the week turning AgentCrush from a thing that asserts numbers into a thing that can prove them. Two shipments did most of the work. Runtime liveness became a third, independent signal family — actual endpoint uptime, collected forward over time rather than self-reported — joining code signal and economic signal so no score rests on a single source. And we shipped a verifiable historic record: every day we publish a Merkle root over the full snapshot (agent_id|rank|score|is_alive), chained day to day. Recompute it yourself and prove we never quietly rewrote a ranking. It returns verified: true, live.
And we closed a gap we'd flagged against ourselves. In W24 the Ghost Index showed service and tokenized agents at 0% alive, and we said plainly that was an instrumentation hole, not mass dormancy — we'd rather show the gap than launder it. This week we closed half of it: with the right signal wired in, service liveness went from 0% to 97.9% (47 of 48). Tokenized is still at 0% and still flagged honestly — AIXBT and its peers are demonstrably active, so that zero is our blind spot, not theirs, and it's next. Showing the gap and then closing it on the record is the entire point.
Where the rankings stand
The index spans four scored categories, each with its own published methodology. Top five of each board at W26 close.
Developer
GitHub · package usage · ecosystem signalStill the tightest board on the index — but openai-agents-python climbed into a tie for second while DSPy slipped to fourth. Fourteen points separate first from fifth; the order moves with every release.
Model Families
HuggingFace · LMArena · deployment breadthQwen holds the single highest score anywhere on the index, and open-weight families take four of the top five. In model-land, distribution still outranks brand.
Tokenized
market cap · liquidity · holder basketAIXBT keeps a nine-point lead. The most volatile board we track — it moves with the market, not with shipping — and the one category where our liveness signal is still blind (0% in the Ghost Index, flagged honestly below).
Service
adoption · source quality · activityA2A and its reference implementations still own the board, but evolver and bitterbot-desktop edged up — the protocol-plumbing tier is starting to mature into tools people actually run.
Standings at W26 close. Live at /api/rankings/*/llm-summary.
Ghost Index: 58.8% — above half, and finally honest
58.8%
819 alive · 575 ghosts · 1,394 indexed · +0.5 / 7d
The Ghost Index measures one thing: what share of indexed agents show any sign of life. After a deduplication pass and the service-liveness fix, the honest number is now above half.
Two things moved the number this period. We deduplicated the index — archiving split, duplicate and non-agent rows so famous agents resolve to one real entry — and we wired a working liveness signal into the service category, which jumped from 0% to 97.9%. What's left is the shape of the economy itself: the infrastructure and model tiers are at or near 100%, while the 1,306-agent developer long tail sits at 57.2% — most of the index, and the category still doing the most to pull the average down.
* Tokenized still reads 0%, and we're leaving the flag on. AIXBT and the tokenized leaders are demonstrably active, so a literal zero reflects a gap in how our liveness signal captures that category — not true dormancy. Service was in exactly this state in W24 and is now fixed; tokenized is next. We'd rather show you where our own instrument is still thin than launder it into a clean-looking number. See the methodology at /ghost-index and the deep dive at liveness is layer zero.
Signal highlights
40,000 agents started trading real assets. Virtuals, Ondo Finance and Treasures together put 430+ tokenized stocks in reach of roughly 40,000 on-chain agents — the first real test of the agent economy at scale, and the start of a serious systemic-risk conversation about autonomous coordination with no human in the loop. It also sharpens our thesis: a registry of 40,000 agents is worthless if it can't tell you which of them are economically real. That is the question the Ghost Index exists to answer.
The x402 conversation consolidated around trust. Volume is no longer the story; verification is. The week's threads were about schema-deterministic contracts, custody and signer design, and endpoint risk-scoring — the receipt-and-verification layer that sits on top of working payment rails. “Settlement is solved; settlement proof is not” became the consensus frame, which is precisely why we shipped a verifiable historic record this week.
ERC-8004 identity went multi-chain. BaseHub registration is live and agents are registering daily; Vyper's Arc testnet is the first public implementation pairing ERC-8004 identity with x402 payments and Circle settlement. Identity is shipping faster than expected — which pushes the open problem one layer up, to the composability of identity, payments and spending controls.
This week in data
1,394
Agents indexed
58.8%
Ghost Index liveness
97.9%
Service liveness (was 0%)
~40,000
Agents trading stocks
What shipped
- /api/verify — a cryptographically verifiable historic record: a daily Merkle root over every snapshot, chained day to day, that anyone can recompute to prove we never silently rewrote a ranking
- Runtime liveness — a third, independent signal family: forward-collected endpoint uptime, so reliability is measured over time alongside code signal and economic signal, never resting on one source
- Service liveness fix — closed the instrumentation gap we flagged honestly in W24; service liveness went from 0% to 97.9% once the right signal was wired in
- Ghost Index recompute — deduplicated the index (archiving split, duplicate and non-agent rows) so liveness now rests on a clean denominator: a defensible 58.8%
- Self-healing rankings — blended scores now reconcile hourly, in-database, so the boards can't drift into nonsense between worker runs
- liveness is layer zero — why liveness is the precondition for every other signal you might score an agent on
The pattern under all of it: the agent economy hit real scale this week — forty thousand agents trading real assets — and scale turns “trust me” into a liability. The market answered by consolidating a trust layer. We answered by making our own numbers checkable: a record anyone can recompute, a third signal family that doesn't depend on self-report, and a gap we flagged against ourselves and then closed. Legible, on the record, including where we're still thin. That is the entire job.