NightBuild #3 — reference sheet
Three harnesses and three loops, each written as a portable Markdown spec. Every entry expands to show the full specification, which you can copy into any platform's session-management files (CLAUDE.md, AGENTS.md, or equivalent).
Harnesses
Coding agent harness
Frontier model designs tasks and security gates; cheaper model executes against PROGRESS.md, halting rather than self-patching a failed gate.
Research agent harness
Frontier model scopes sub-questions and source gates; cheaper model searches and logs claims to FINDINGS.md, escalating contradictions instead of resolving them.
Ops / deployment agent harness
Frontier model sets promotion gates and rollback triggers; cheaper model runs staged rollout against STATE.md, with production always halting for human approval.
Loops
Test-driven fix loop
Companion to the coding harness. Writes the failing test first, iterates fixes against the full suite, and halts on stalled progress instead of burning retries.
Source triangulation loop
Companion to the research harness. Varies search angle per iteration, dedupes by publisher, and routes agreement checks to a cheaper model but contradictions to the frontier model.
Canary rollout loop
Companion to the ops harness. Ramps traffic multiplicatively with fixed observation windows; automatic promotion is allowed, automatic rollback is not.