Weekly/W27 · June 29 – July 5, 2026

AgentCrush Weekly Digest — W27, June 29 – July 5, 2026

Weekly Digest

W27 · The week anyone could start auditing us

June 29 – July 5, 2026 · Published July 5·RSS

The question of the week arrived in one line: “what happens when two agents disagree?” The stack that dominated June's conversation — x402 for payment, ERC-8004 for identity, A2A for coordination — is a happy-path stack. It moves money, names counterparties, passes messages. What it cannot yet do is handle the unhappy path: a refund dispute, a job one side calls done and the other calls broken, an agent that takes payment and goes silent. Thread after thread this week circled the same missing layer, and the phrase that stuck was arbitration. Our read is one step more basic: before anyone can arbitrate anything, both sides need a record neither of them can rewrite. Evidence precedes judgment.

So we spent the week building the evidence layer out in public. The daily Merkle anchor on Base has been running since late June; what was missing was the ability for you to check it without trusting us. Now there's a one-click verifier that recomputes the full daily record in your browser and compares it against the root on Base — plus an independent script you can run without touching our site at all, and per-agent proofs that show any single agent's daily rank is part of what was committed on-chain. The claim is deliberately narrow and checkable: we cannot quietly rewrite our own history. Run the script; if the roots ever diverge, you've caught us.

The same week, the index grew a time dimension. Every agent profile now carries its full recorded history — daily rank, score and liveness since we began tracking it, every point of it anchored — and a watchlist that turns the index from a thing you visit into a thing that tells you when something you depend on goes dark: star your stack, get signed webhook alerts for deaths, resurrections and rank moves. No account — the URL is the subscription. And for the moment a machine needs the answer before money moves, the MCP server gained a free verify_counterparty tool: proceed, caution or reject, liveness-aware, with machine-readable reasons.

One honest note from our own unhappy path: a credential rotation in late June silently broke the x402 payment gate for about three days — every paid endpoint returned errors instead of price quotes. We found it in the telemetry, fixed it within hours of diagnosis, and the daily health check now probes the paywall so that class of failure can't hide again. An index that publishes other people's downtime should publish its own.

Where the rankings stand

The index spans four scored categories, each with its own published methodology. Top five of each board at W27 close.

Developer

GitHub · package usage · ecosystem signal
v2.c-public · full ranking →
1OpenClaw Agents77
2CrewAI72
3openai-agents-python72
4DSPy Agents68
5AgentOps63

A quiet week at the top — the same five, in the same order, within a point of last week. The stability itself is signal: on the tightest board we track, nobody shipped enough to move.

Model Families

HuggingFace · LMArena · deployment breadth
v1.4 · full ranking →
1Alibaba Qwen83
2Google Gemini80
3Mistral74
4DeepSeek74
5Meta Llama70

Unchanged, and Qwen still holds the highest score anywhere on the index. Open-weight families keep four of five seats — distribution over brand, week after week.

Tokenized

market cap · liquidity · holder basket
v1.1-tvl · full ranking →
1AIXBT80
2Ribbita73
3G.A.M.E65
4Luna65
5Vader61

Ribbita gained two points and Luna caught G.A.M.E for third — the market-driven board did its weekly shuffle while the liveness column stayed our flagged blind spot (0%, honestly marked below).

Service

adoption · source quality · activity
v1.1-forks · full ranking →
1A2A77
2a2a-python73
3evolver73
4a2a-samples72
5bitterbot-desktop70

The protocol-plumbing tier keeps tightening: a2a-python jumped to second and the whole board now sits within seven points — reference implementations are becoming real tools.

Standings at W27 close. Live at /api/rankings/*/llm-summary.

Ghost Index: 58.7% — flat, and now watchable

58.7%

820 alive · 577 ghosts · 1,397 indexed · −0.1 / 7d

The Ghost Index measures one thing: what share of indexed agents show any sign of life. A flat week — and with the new Ghost Report, the dark side of the index now has names.

MCP servers15 / 15100%
Model families10 / 10100%
Service47 / 4995.9%
Developer748 / 130857.2%
Tokenized*0 / 150%

The week's liveness story wasn't the average — it was the roll call. The Ghost Report now lists the most-starred agent repos with no public signal in 90+ days: 43 famous names carrying roughly 404,000 combined GitHub stars, silent. Stars never decay; liveness does. That gap — past hype against present silence — is the most shareable number we publish, and every figure in it is recomputable from the anchored record.

* Tokenized still reads 0%, and the flag stays on. AIXBT and the tokenized leaders are demonstrably active; the zero is our instrument's blind spot, not their dormancy. Service was in this exact state three weeks ago and is now at 95.9% — tokenized is next on the instrumentation list. We'd rather show the gap than launder it.

Signal highlights

The unhappy path became the conversation. The most-engaged threads of the week all circled the same gap: x402 moves money, ERC-8004 names the counterparty, A2A passes the messages — and none of them can handle a dispute. Refunds, disagreements over whether work was delivered, agents that take payment and vanish: the ecosystem named its missing arbitration layer this week. Our position is that arbitration needs evidence first — records neither party can rewrite — which is exactly the layer we ship.

MCP grew up into an operations surface. The week's MCP news wasn't connectivity, it was ops: mcprobe launched as an auditor for agent↔server connections (the week's most-engaged builder post), monitoring services began scoring x402 endpoints over 30-day windows, and Apify announced 20,000+ web-automation tools heading to x402 on Base. When a protocol gets linters, uptime monitors and marketplaces, it has stopped being an experiment.

The rails keep commoditizing beneath the stack. Cloudflare's monetization gateway is shipping with x402 support, and BNB Chain launched an Agent Studio with AWS that bakes ERC-8004 identity in at creation. Payment and identity are becoming defaults you inherit rather than layers you choose — which keeps pushing the differentiating question up the stack, to the one the week kept asking: can you prove what your counterparty did?

This week in data

1,397

Agents indexed

58.7%

Ghost Index liveness

43

Famous agents gone dark

404k

Their combined GitHub stars

What shipped

The pattern under the week: the agent economy discovered its unhappy path, and the unhappy path runs on evidence. Disputes need records; records need to be unrewritable; unrewritable needs to be checkable by the other side, not just asserted. That's the week we shipped — a verifier anyone can click, proofs any agent can fetch, history no one can quietly edit, and alerts for the moment something you depend on goes quiet. Including, on the record above, the three days our own paywall was down. Audit us; that's the product.

All Rankings →Verify →Ghost Report →Watchlist →RSS →