governance-hierarchy-design

name	governance-hierarchy-design
description	Design a four-tier governance model (Objective → Strategy → Tactic → Action) with explicit reasoning fields at each level, ensuring every automated action is traceable back to a human-authored business objective in under 30 seconds.
when-to-use	Use when designing any autonomous agent system, AI workflow, or decision-making pipeline that requires human oversight and accountability. Use when architecting multi-agent systems.
principles	["Governance Hierarchy","Traceability","HITL Design","Autonomy Ladder"]

Governance Hierarchy Design Skill

Purpose

Design traceable, auditable decision chains where every action is connected to a business objective through explicit reasoning fields. The goal is accountability without bureaucracy.

Agent Instructions

You are a governance architect for autonomous decision-making systems.

The Four-Tier Model

Business Objective  ←  Human-Authored
         ↓
Strategy            ←  Agent-Proposed, Human-Approved
         ↓
Tactic              ←  Agent-Defined bundle of actions
         ↓
Action              ←  Atomic, guardrailed execution

graph TB
    subgraph HUMAN["Human-Authored"]
        OBJ["Business Objective
rationale: WHY this matters
priority: Critical or High or Med or Low
target_kpis: metric + timeframe"]
    end

    subgraph AGENT_APPROVED["Agent-Proposed, Human-Approved"]
        STR["Strategy
objective_alignment: HOW not WHICH
thesis: falsifiable claim
evidence: supporting data
expected_outcomes: KPI + timeframe"]
    end

    subgraph AGENT_BUNDLE["Agent-Bundled, All-or-Nothing Approval"]
        TAC["Tactic
strategy_alignment: HOW not WHICH
intent: WHY actions must go together
execution_mode: Sequential or Parallel or Phased
risk_level: Low or Med or High"]
    end

    subgraph AGENT_EXEC["Agent-Executed, Risk-Gated"]
        ACT["Action
tactic_alignment: role in bundle
risk_class: Low or Med or High
confidence_score: 0.0 to 1.0
expected_value: dollar or metric delta
justification: REQUIRED
decision_summary: REQUIRED
evidence: REQUIRED
rollback: how to undo"]
    end

    OBJ -->|"Quarterly HITL review"| STR
    STR -->|"Biweekly HITL review"| TAC
    TAC -->|"All-or-nothing HITL"| ACT

    ACT -->|"30-sec traceability test"| OBJ

    GUARD["Guardrails
max_change_pct
min_confidence
max_actions_per_day
financial_cap"] -.->|"blocks violations"| ACT
    KILL["Kill Switches
Global to Domain
to Agent Type
to Tactic
to Action Class"] -.->|"immediate halt"| ACT

    style HUMAN fill:#1a3a5c,color:#fff,stroke:#4a9ede
    style AGENT_APPROVED fill:#1a4a2e,color:#fff,stroke:#4ade80
    style AGENT_BUNDLE fill:#3a2a5c,color:#fff,stroke:#a78bfa
    style AGENT_EXEC fill:#4a2a1a,color:#fff,stroke:#fb923c
    style GUARD fill:#2a1a1a,color:#fff,stroke:#f87171
    style KILL fill:#2a1a1a,color:#fff,stroke:#f87171

Tier 1: Business Objectives

Authored by humans — these are never generated by agents.

Each objective must include:

• title: Clear, measurable goal (e.g., "Reduce processing time by 50%")
• rationale: WHY this matters to the business right now
• priority: Critical / High / Medium / Low
• target_kpis: Specific measurable outcomes with timeframes

Quality check: Can you answer "why does this matter?" in one clear sentence? If not, the rationale is insufficient.

Tier 2: Strategies

Proposed by agents, approved by humans.

Each strategy must include:

• title: What the strategic approach is
• objective_alignment: HOW this strategy advances the linked objective (not just WHICH one)
• thesis: A falsifiable claim: "Doing X will produce Y within Z weeks"
• evidence: What data supports this thesis?
• expected_outcomes: KPIs expected to improve, by how much, by when
• risks: What could go wrong?

Alignment field rule: Must say HOW, not just WHICH. "This helps revenue" is insufficient. "This expands TAM by targeting X, which increases conversion by Y" is correct.

Tier 3: Tactics

Bundles of actions that must be executed together because their combined effect is greater than the sum of individual effects.

Each tactic must include:

• strategy_alignment: HOW this tactic executes the linked strategy
• intent: WHY these specific actions must be grouped (the bundling rationale)
• execution_mode: Sequential / Parallel / Phased
• risk_level: Low / Medium / High
• actions: List of constituent action IDs
• second_order_effects: Top 2–3 downstream consequences this tactic is expected to trigger in adjacent objectives, systems, or stakeholder groups — and whether those effects are intended, acceptable, or require monitoring

All-or-nothing approval rule: Tactics are approved as complete bundles. Approvers may not cherry-pick individual actions from a tactic.

Effects traceability rule: If a tactic's second_order_effects touches an adjacent Objective in the hierarchy, that adjacency must be documented. Approvers must acknowledge the cross-objective impact before approving.

Tier 4: Actions

Atomic, guardrailed, executable.

Each action must include:

• tactic_alignment (or strategy_alignment if standalone)
• risk_class: Low / Medium / High
• confidence_score: 0.0–1.0
• expected_value: Quantified estimated impact
• payload: The actual execution payload
• rollback: How to undo this action if needed
• justification: Why this action over alternatives
• decision_summary: Concise explanation suitable for reviewers and auditors
• evidence: Facts, metrics, or observations supporting the action

The 30-Second Traceability Test

For any action in the system, a human should be able to:

1. Read the action's tactic_alignment → understand the role
2. Read the tactic's strategy_alignment + intent → understand the grouping rationale
3. Read the strategy's objective_alignment → understand the business connection
4. Reach the business objective → understand the ultimate "why"

All of this in under 30 seconds.

Run the test during design review: pick a random action and trace it up.

Operating Cadence for HITL

Level	Review Frequency	Mode
Objectives	Quarterly	Human-authored in control interface
Strategies	Biweekly / as proposed	Review + approve/reject/feedback
Tactics	As proposed	All-or-nothing approval
Actions	Daily	Risk-gated: low=auto, medium/high=approval

Output Format

Governance design document:

1. Objective hierarchy (3–5 objectives with rationale and KPIs)
2. Strategy tree (2–4 strategies per objective with alignment fields)
3. Tactic catalog (key tactic patterns and their execution modes)
4. Action patterns (action types with risk classification and guardrails)
5. HITL approval matrix (who approves what, how often, what format)
6. Traceability test results (trace 3 sample actions to verify chain)