/harness · the system that produces the system

The harness is the product.

In an AI-native company the highest-leverage work is not designing one screen — it's continually refining the harness that produces ten thousand. That's how a small team ships at SageOx's 2026–2027 scale.

this page the layers · the cross-cutting principles · pointer to the full HARNESS.md

One hour polishing one screen vs. ten thousand. Polish the system, not the artifact. Every hour improving the harness compounds across every surface that ships after.

— DESIGN.md §0 Rule 4 · the harness is the product

Who you are → where to start

🎨 Design engineer

Refines the harness — tokens, lint, DESIGN.md, agent context. Also ships features end-to-end.

⚙️ Product engineer

Ships features in consuming repos. Consumes tokens + components from this repo; has commit access if the harness has a gap.

🎯 Product owner

Shapes intent, density, voice, microcopy. Ships customer-facing copy directly. (Renamed from PM.)

Persona names describe job-shape, not handoff lines. Everyone here ships production code.

The layers

15 · grouped 4 ways

The harness isn't one thing — it's a coordinated set of layers that together let any agent produce on-brand, accessible, themed UI without further prompting. Missing one creates drift; missing several means the system isn't a system. Full descriptions and source paths in HARNESS.md Part 2.

Foundation what exists

repoA dedicated sageox/sageox-design repo. Not buried in product code.
spec · canonicalDESIGN.md — the canonical guidance. Best-effort, not enforced; the lint/CI/eval layers below are what make it stick.
spec · mirror.claude/rules/design.md per consuming repo — lightweight, agent-facing.
tokensColor, type, spacing, motion. Canonical here; mirrored into each surface repo.
componentsPre-built primitives with great defaults, per surface.
themesLight + dark + calm. Mandatory on every surface; not optional polish.
brand contractsPixel-exact assets that can't be regenerated — wordmark, icon, mascot.

Enforcement what catches drift

lint / CIContrast checks, token-usage lint, motion budgets, theme coverage. Fails loud.
evalSpot audits of agent-produced surfaces. The feedback signal driving Loop A.

Agent infrastructure what agents reach for

agent contextCLAUDE.md + per-repo equivalents — framing that tells agents how to read the spec.
subagentsSpecialist review lenses: ui-designer, hci-expert, perceptual-ux, accessibility-expert, agent-ux.
catalogsLive agent-iterable component catalog on every surface — required by DESIGN.md §11 Rule 7.

Distribution how it propagates

sync scriptscripts/sync-design-system.sh in every consuming repo. The system you can't pull goes stale.
portalsageox-design.netlify.app — readable by humans and agents alike.
studioproposals/ — dated mockups stress-testing changes before codification.

Six cross-cutting principles

true everywhere, or true nowhere

Properties of the harness that don't fit in a single layer — they cut across the whole system. If any of these is true in some surfaces and not others, the harness fails the surfaces where it isn't true.

Two-layer spec

Canonical DESIGN.md + lightweight per-repo mirror, regenerated by sync, pinned by commit SHA. Agents shouldn't traverse out to canonical for every read.

Theme matrix is first-class

Light + dark + calm. The catalog's theme switcher is how agents verify their output covers all states. Not optional polish.

Distribution or drift

Every consuming repo ships a one-command sync from sageox-design. Pinned snapshot + optional --check CI gate. Else stale.

Many lenses, not one human

Five specialist subagents review on demand. Five fast review passes by lens beats one slow review by an overloaded human.

Agent UX ≠ human UX

Agents are peer systems consuming our UI. Stable semantic class names, predictable JSON output, idempotent actions, designs that survive being copied as text.

Designers are coaches

~90% of the work is writing principles, delegating to subagents, and auditing a swarm — not pixel-tweaking. Best hires have led design teams or run agent swarms at scale.