/decisions · 2026-05-12 · ops · ray-capability
Proposed: a 4-agent pipeline (spec author → test author → code author → critic) for Ray skill builds, gated by founder spec sign-off, addressing the tests-rubber-stamp-code anti-pattern from this morning's MAC testing conversation.
This morning we worked through Garry Tan's "Agent Complexity Ratchet" testing framework. The pushback you raised: when the same person writes the spec AND the test AND the code, tests rubber-stamp the code instead of validating against ground truth.
The proposal: separate the Ray skill build-out into 4 sub-agent roles, each reading from upstream artifacts only:
This builds on the existing queued Pre-Launch Customer Simulator skill candidate (from /improve 2026-04-20). If approved, I file a Notion ticket with full scope + acceptance criteria; build is queued for a future cycle.
File the Notion ticket. Add any scope adjustments or specific requirements you want baked in.
Approve — file the ticket + sendSkip this for now. Add a one-line reason so the kill is on the record.
Archive — not now + sendKeep part of the proposal, defer the rest. Use the two fields to describe each.
Split + send