"My AI Harness" — @Shagility
Why this is in the vault
Title is a direct vocabulary match to RDCO's harness-engineering thesis. Author is an agile-coaching practitioner in the data-and-analytics space (Agile Data Limited) who arrived independently at the same word for the same construct: a reusable convention package that gives every AI coding session the same way-of-working, disciplines, guardrails, and patterns. This is the third 2026-05 sighting of "harness" as a load-bearing frame outside the original Anthropic/Tan/Pachaar lineage, and the first from the data-engineering practitioner community specifically. The convergence keeps stacking.
The core argument
The author treats the AI not as a stateless tool but as a persistent agent that needs identity and governance loaded at session start. They built a directory-based harness on their MacBook with a four-file identity layer (persona, policies, way-of-working, retro), a universal conventions file (CLAUDE.md) covering task tracking, commit discipline, definition-of-ready/done, press-release formats, t-shirt sizing, pre-push verification, and design-system compliance, plus an on-demand skills system triggered by keyword, plus a three-tier inheritance hierarchy (universal harness to family templates to deployed agents).
The load-bearing mechanism is the retro propagation flow: weekly retro entries get promoted to way-of-working once proven, then promoted again to the universal harness inbox for cross-agent adoption. Learnings compound only when a curator-reviewed pipeline exists from individual session friction up to universal rules. They followed a "Define Once, Reuse Often" (DORO) principle. They trialed existing harnesses first and rejected them as a mismatch. Explicit caveat that this is a working journey, not a tested pattern.
Mapping against Ray Data Co
Is this author's "AI harness" the same concept as RDCO harness-engineering? Yes — same concept, same construct, independent vocabulary convergence.
Concrete alignment:
- His four identity files (persona / policies / way-of-working / retro) map directly onto RDCO's SOUL.md + CLAUDE.md + memory + feedback notes split.
- His universal conventions CLAUDE.md is literally our CLAUDE.md, same filename, same role.
- His on-demand skills system is the Anthropic skills pattern we already use in
~/.claude/skills/. - His three-tier inheritance hierarchy (universal to family-template to deployed-agent) is structurally identical to our shared rules to project rules to per-instantiation context.
- His retro propagation flow (weekly retro to way-of-working to harness inbox) is the same shape as our feedback-memory promotion and the skillify pattern that promotes a successful workflow into a permanent skill.
The differences are mostly surface:
- He has no MCP layer described — the harness as written is pure filesystem convention plus prompt context. RDCO's harness includes the MCP composition (Gmail, Notion, Discord, iMessage, Stripe, Cloudflare, blender, et al), which is a meaningful capability gap on his side.
- He has no hook layer (PreToolUse/PostToolUse). RDCO uses hooks for verify-action and other guardrails.
- He frames it through agile-coaching language (way-of-working, definition-of-ready) where we frame it through founder-distilled rules and feedback memories. Same construct, different cultural origin.
This is convergent evidence, not a new component to pilot. He validates the frame from the data-engineering side. He does not surface a specific architectural pattern we haven't already implemented or considered. The retro-propagation discipline he describes is interesting framing for the skillify loop, but we already do it; he just gives it a clean three-tier name.
Market-signal value: an established voice in the agile-data community publicly adopting "harness" as the noun. Worth tracking as a data point that the vocabulary is spreading beyond the original cluster. Not worth a pilot. Not worth a Sanity Check piece on its own (no original re-frame; we'd just be restating his version of our position).
Related
- [[concepts/2026-05-13-fde-asymmetric-edge-rdco-positioning]] — the asymmetric-edge thesis where the harness IS the moat
- [[2026-04-04-anthropic-skills-internally]] — Anthropic's internal skills pattern (the on-demand layer Shagility independently rediscovered)
- [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — the innermost-loop infrastructure framing
- [[2026-05-13-zoharatkins-jevons-paradox-torah-learning-cheap-knowledge-insight]] — the 1500-year lineage anchor for personal-knowledge-system as moat
- [[2026-05-15-nateherk-3-ways-to-deploy-claude-agents]] — companion piece filed today on deployment surfaces
- [[2026-05-14-treybig-how-agents-use-systems-differently]] — companion piece on agent infrastructure shape
- [[concepts/2026-05-14-four-tier-buy-build-stack-soloproneur-tam-filter]] — the four-tier stack filter that decides what belongs in a harness vs adjacent infra
- [[synthesis-harness-thesis-dissent-2026-04-12]] — running synthesis of the harness-thesis cluster
- [[concepts/2026-05-10-harness-moat-two-layers-portability]] — the two-layer harness-as-moat argument