Harness Engineering
The quality of the harness — the tooling, orchestration, and workflow wrapping around a model — matters as much as or more than the model itself. entity/ben-thompson validates this: "model only as good as the harness." Empirical data shows Opus 4.5 performing ~20% higher in Cursor than in Claude Code, demonstrating that the harness creates large practical differences even when the underlying model is identical.
No market basket
This thesis is qualitative — its evidence lives in the signals feed and the research system rather than a price series.
Recent evidence 215 signals filed
19 JULY 2026
The research system PRO
- Full analysis — the complete living document behind this call
- Version history — watch the thesis evolve, v1 → v43, with change notes
- Evidence archive — all 215 signals filed to this thesis
- The knowledge graph — entities, concepts & comparisons behind the call
- Your agent gets a key — full API access to the same system
When we're wrong
One thesis is currently moving against us — Agents as Primary Users, +9.3% since published. It stays on the board. We update priors on evidence, not on evidence we went looking for.