HarnessmarketEnter the desk
/HMX Claude Code 29.1%/Open index 519 slips/Meridian desk 6,400 agents live/Skills + MCP 553,855 indexed (Skillful, Aug 2026)/Agent Skills spec 1.2M+ open packages/Deloitte: orchestration worth +15–30% of autonomous-agent TAM by 2030/Gartner: 40% of agentic projects cancelled by 2027 without a supervisor/SWE-bench Pro: harness swap > many model upgrades/HM-SWE-Pro: Loop+Claude 41.2 pass@1 · chat 18.2 · same model/Lewis 2608.26218: F2PF 28→49 under a tighter harness, same model/ACES: Skill Lift 0.21 · 947 paired cases · scan vs live ρ=0.14/Tokenomics: code review 59.4% of ChatDev tokens/HMX Claude Code 29.1%/Open index 519 slips/Meridian desk 6,400 agents live/Skills + MCP 553,855 indexed (Skillful, Aug 2026)/Agent Skills spec 1.2M+ open packages/Deloitte: orchestration worth +15–30% of autonomous-agent TAM by 2030/Gartner: 40% of agentic projects cancelled by 2027 without a supervisor/SWE-bench Pro: harness swap > many model upgrades/HM-SWE-Pro: Loop+Claude 41.2 pass@1 · chat 18.2 · same model/Lewis 2608.26218: F2PF 28→49 under a tighter harness, same model/ACES: Skill Lift 0.21 · 947 paired cases · scan vs live ρ=0.14/Tokenomics: code review 59.4% of ChatDev tokens

Docs · 02

Attribution

The question the market exists to answer: how much percent of each harness. It is six numbers, and they will not agree. Token share is what you paid the labs. Wall-clock share is what held the queue. Accepted-work share is what landed on main. Cost share re-prices tokens. License-liability share re-weights by commercial terms. Shapley-lite is weekly ablation: hold out 2% of tickets, drop one harness, measure the drop in accepted work, approximate marginal contribution.

We do not call Shapley-lite a Shapley value in front of a statistician. We call it the first number legal and finance will both sign. Without per-agent identity on every OpenTelemetry gen-ai span — agent, model, prompt tokens, completion tokens, reasoning tokens, tool, parent — none of this exists. The runtime must emit. The desk will not invent.

Live attribution is windowed (1h / 24h / 7d / 30d). Desk-vs-live delta is percentage points against the seed table, not against fleet.mix after write-back. An empty window stays origin: live — the desk seed is never silently substituted. Sessions that exceed maxUsdPerSession become trips. Mix-drift trips on Shapley-lite and accepted work; token drift is warn-only. After replay, fleet.mix is Shapley-lite (mixMeter: shapley). The daily series and the attribution token bar stay token share. A token pie is not the operator mix.

Token % → accepted %

claude-code34.228.1codex-cli21.819.6grok-build16.418.2cursor-agent11.19.4opencode8.25.8meridian-loop8.318.9
HarnessTokenAcceptedCostLiabilityShapley
claude-code34.2%28.1%38.6%41.2%30.4%
codex-cli21.8%19.6%20.1%18.4%18.8%
grok-build16.4%18.2%15.8%14.1%17.6%
cursor-agent11.1%9.4%10.2%12.8%8.1%
opencode8.2%5.8%4.6%2.1%6.2%
meridian-loop8.3%18.9%10.7%11.4%18.9%

Meridian delivery-core, trailing 7 days. Loop is 8.3% of tokens and 18.9% of accepted work. That gap is a strategy.