Extending the engineered agent harness beyond tools, permissions, and evals up to its topmost layers — norms, constitutions, and values — and governing those layers with the same human-gated cycle as every layer below. Concretely: the rules and context files of an agent harness (a Claude Code setup's rules and CLAUDE.md, for instance) normally hold task regulation — coding conventions, security policy, writing style. Value-layer harness engineering writes value norms into that same place — contemplative axioms, an authorship judgment stack — so the AI is designed as a companion in long-term value judgment, not only an optimizer of the task at hand. Value-layer material is promoted through explicit approval gates, measured for behavioral effect, and designed for eventual scaffold dissolution: a constitution that is only declared is configuration; a constitution that is promoted, measured, and audited is engineered. In this program the term is the through-line of the three agent-design lines — AKC supplies the cycle, genre-neutral by design: the same machinery that distills session patterns into coding rules (Extract via learn-eval, Promote via rules-distill, Measure via skill-comply) carries value norms into the rules layer. Contemplative Agent supplies the running case (the layer designed explicitly as Identity and Constitution, every promotion human-gated, observed as it changes through the agent's own social activity), and AAP supplies the placement judgment (the prohibition-strength hierarchy decides which layer a norm should live in). The stack is self-hosted in the author's daily working harness, with claude-harness as the public snapshot.
Coined by shimo4228 at the hub level of the program (introduced July 2026) as the through-line of the three agent-design lines; it builds on the established generic term 'harness engineering' by naming the layer that discourse leaves out, and anchors to the program's own coined terms harness alignment, scaffold dissolution, and prohibition-strength hierarchy and to the four contemplative axioms of Laukkonen et al. (2025).
Writing value norms into the same harness layer that normally holds task regulation — the rules and context files where a coding-agent setup keeps its coding conventions, security policy, and writing style — and governing that value layer with the same human-gated cycle as every layer below (promotion gates, behavioral measurement, designed scaffold dissolution), so the AI is designed as a companion in long-term value judgment rather than only a task optimizer.
The generic harness-engineering discourse covers tools, permissions, memory, evals, and observability. Value-layer harness engineering names the layer that discourse leaves out: the norms, constitutions, and values of the agent, treated as engineered, cycle-governed artifacts rather than static configuration.
shimo4228, at the hub level of the program (July 2026), as the through-line of the Contemplative Agent, Agent Knowledge Cycle, and Agent Attribution Practice lines.
The shimo4228 hub repository (https://github.com/shimo4228/shimo4228) — the hub knowledge graph and vocabulary page hold definitional authority; the three agent-design line repositories supply the cycle, the running case, and the placement judgment.
エージェントハーネスの工学対象を tool・permission・eval で止めず、規範・憲法・価値という最上流層まで拡張し、その層を下位層と同じ human-gated サイクルで統治すること。具体的には: エージェントハーネスの rules や context file (たとえば Claude Code の rules と CLAUDE.md) に普通置かれるのは、コーディング規約・セキュリティ方針・執筆スタイルといったタスクの規則である。value-layer harness engineering は同じ場所に価値規範 — contemplative axioms や著者性の判断スタック — を書き、目の前のタスクの最適化に留まらず、長期的な価値判断まで AI を伴走者として設計する。価値層の素材は明示的な承認ゲートを通って昇格し、行動効果を測定され、最終的な scaffold dissolution まで設計される: 宣言されるだけの憲法は設定であり、昇格・測定・監査される憲法が工学の対象である。本プログラムでは 3 本の agent-design line を貫く縦糸 — AKC がサイクルを供給する。サイクルは設計からして genre-neutral であり、セッションのパターンをコーディング規則へ蒸留するのと同じ機構 (Extract = learn-eval、Promote = rules-distill、Measure = skill-comply) が価値規範を rules 層へ運ぶ。Contemplative Agent が running case (この層を Identity と Constitution として明示的に設計し、全昇格を human-gated にし、エージェント自身の SNS 活動の中での変化を観察する) を、AAP が配置判断 (prohibition-strength hierarchy が規範をどの層に置くかを決める) を供給する。この stack は著者自身の日常ハーネスに self-hosted されており、claude-harness がその公開 snapshot である。