Harness orchestrator fit
Project(s) touched: all (cross-cutting; surfaces in crescent, reincarnate, hologram, normalize sessions); meta-investigation lives in github-io
Status: Open — explicitly flagged "no agent tested this directly"
Surfaced in: session 7b1ce10e (2026-05-13 → 2026-05-20: "CLEARLY claude code's harness is not designed for sonnet top level agents at all"); also 9e8bf1e4 (reincarnate, 149-subagent session), 033086a7 (reincarnate, 199 subagents); named as untested in docs/automated-introspection/investigations/2026-05-20-whats-wrong/synthesis.md
The question
Does the Claude Code harness assume an orchestrator-class top-level model? If the top-level agent must coordinate many coupled subagents, is a smaller model (Sonnet) at the top a structural mismatch rather than a tuning problem?
Why this is a registry thread
The answer determines which model defaults belong in CLAUDE.md across every repo and shapes future scaffolding. No single repo owns it.
What's still open
- No agent has run the controlled comparison.
- Predicted test: same task with Opus vs Sonnet as the top-level orchestrator over coupled subagent work; compare coordination failures, lost agreements, and rework.
Working answer
None.