Archive/State of the Index — September 2026
Monthly Report
State of the Index — September 2026
The Ghost Index opened September at 54.1%, dipped to a low of 52.7% on September 12, then stabilized in a narrow 52.9–53.4% band through the second half of the month, closing at 52.9% on September 30 — a net decline of 1.2 percentage points. The indexed base grew by 12 agents (1,435 → 1,447) while alive agents fell by 11 (777 → 766). The pattern follows the same logic it has all year: the index onboards agents faster than their activity signals corroborate, and liveness settles gradually as the 30-day window fills. Service held at 78.9% (60 of 76); Developer stayed just above half at 51.2% (681 of 1,331).
September's headline board move was on the Developer ranking: openclaw climbed from #3 to #1 (71→74.5), while CrewAI fell from #1 to #3 (74→71.6). OpenAI Agents Python held at #2 with a modest score gain (72→73.4). DSPy Agents dropped out of the top 5; a2a-python entered at #5 at 67.1 — the first appearance of an A2A-protocol agent on the Developer board. The Model Families board was static: same five families, same order, same scores as August. On Tokenized, AIXBT strengthened its lead (+3 to 83) and Ribbita gained 2 points (75). Voice of the Gods (ADM) displaced Vader at #5 with a 61. The Service board saw a2a-samples return to #4 (72) as agent-teams-ai dropped off.
The evidence-ranked tier reached 212 agents at September close — 21 more than August's 191, continuing the steady expansion pace that has held since May. The archive compounded to 178,477 daily snapshots as of October 1, adding 43,210 new records in September alone. Every snapshot is Merkle-anchored to Base at 04:30 UTC the following morning and independently recomputable via /oracle. September marks month six of an unbroken nightly snapshot run since April 2026.
All four boards at September close
Live from /api/rankings/*/llm-summary as of October 1, 2026.
Developer
GitHub · package usage · ecosystem signalThe month's headline move: openclaw climbed from #3 (71) to #1 (74.5) while CrewAI fell from #1 (74) to #3 (71.6). OpenAI Agents Python held at #2 with a slight score gain (72→73.4). DSPy Agents dropped out of the top 5; a2a-python entered at #5 (67.1). Google ADK Python held #4 at 67 throughout.
Model Families
HuggingFace · LMArena · deployment breadthSame five families, same order, same scores as August. Qwen's derivative count advantage (1,046 fine-tuned variants tracked) and Gemini's deployment breadth lead (145 tracked deployments) held the top two positions unchanged. No methodology changes in September.
Tokenized
market cap · liquidity · TVLAIXBT gained 3 points to reach 83 (#1 since the board launched). Ribbita gained 2 points (73→75). G.A.M.E and Luna held. Voice of the Gods (ADM) entered at #5 with 61, displacing Vader. The liveness column still reads 0% — a standing instrument gap: the probe listens on HTTP while these agents live on-chain.
Service
adoption · source quality · activityThe top three held unchanged. a2a-samples returned to #4 (72) after dropping out in August; agent-teams-ai (which entered last month at #5) fell off. bitterbot-desktop held its score at 70 but slipped to #5. The A2A protocol cluster still occupies the top four; 7 points separate #1 from #5.
Ghost Index: 52.9% — closed 1.2 points below August
52.9%
766 alive · 681 ghosts · 1,447 indexed · Sep 1 open: 54.1%
−1.2 points across September. Dipped to 52.7% on Sep 12, then held a 52.9–53.4% range through the second half. Every number here is independently recomputable via /oracle.
* Two standing flags. MCP's 100% reflects a selection effect — we track actively maintained servers. Tokenized's 0% is an instrument gap: the probe listens on HTTP while those agents live on-chain. The fix is in progress; when the number moves it will move because the measurement improved, stated in those words.
September 2026 in data
1,447
Agents indexed
52.9%
Ghost Index liveness
178,477
Snapshots archived
212
Evidence-ranked (Sep close)
The index runs autonomously — daily snapshots at 02:00 UTC, nightly liveness scoring at 23:50 UTC, Sunday ranking runs, and Merkle anchoring to Base at 04:30 UTC. Every number here is live and independently recomputable via /oracle. The archive is the product: each nightly snapshot is a timestamped record of the agent economy that cannot be backfilled. The data is available machine-readable at /api/ghost-index/v1, /api/rankings/*/llm-summary, and /api/agent-economy/llm-summary.