How much percent of each harness

The question in the original brief is the product: as agents scale, how much percent of each harness. It looks like a pie chart. It is four pie charts, and they disagree.

Token share answers 'what did we pay the labs'. On Meridian delivery-core, Claude Code is 34.2% of tokens and 38.6% of dollars — the model is dearer. Wall-clock share answers 'what held the queue'; Cursor looks larger here because the IDE-resident workers wait on humans. Accepted-work share answers 'what landed on main'; Meridian Loop is 8.3% of tokens and 18.9% of accepted merges, because it is the promotion harness. A token pie would have you fire it.

Shapley-lite is the weekly ablation: hold out 2% of tickets, drop one harness, measure the drop in accepted work, approximate marginal contribution. It is not a true Shapley value and we do not call it one in front of a statistician. It is the first number legal and finance will both sign. License liability re-weights the mix by commercial terms — Claude Code's proprietary seat agreement carries more residual risk than OpenCode's MIT — which is why liability share and token share come apart.

The trace layer is not optional. IBM, Datadog, Langfuse, Helicone, AgentOps, and the OTel gen-ai semantic conventions all say some version of the same thing: every span needs agent identity, model, prompt tokens, completion tokens, reasoning tokens, tool name, and parent. Harnessmarket will not invent what the runtime did not emit. Grok Build 0.9.2 exporting attribution spans is a more important release than a new personality.

Without per-agent attribution, 35–50% cost overruns are not a mystery. They are a missing primary key. The advance on this market is Helix's Shapley-lite pack. The desk you are looking at is the operator's view of that pack after seven days.