agentic-middle-manager（Agent 型中层管理者）

The Management Trinity

Overview

The management-trinity skill is a unified orchestrator that addresses the fundamental limitations of AI agents by unbundling the traditional manager role into a distributed protocol. It synthesizes six critical gap-bridging capabilities into a single, self-improving operating system for the AI era.

Use this skill when:

• Architecting an "agentic organization" or deploying autonomous workflows at scale.
• Transitioning from traditional hierarchical management to a "Player-Coach" or DRI (Directly Responsible Individual) model.
• Designing governance frameworks for high-stakes AI decision-making.
• Addressing systemic issues of trust, burnout, or accountability diffusion in AI-heavy teams.

The Paradigm Shift: From Role to Protocol

When viewed from first principles, the limitations of AI agents are not isolated technical bugs; they are symptoms of a phase transition in organizational design. This skill operationalizes six fundamental paradigm shifts:

1. From "Manager as Role" to "Management as Protocol": Management is no longer a job title held by a single human. It is a distributed protocol where AI handles *Routing* (information logistics), while humans handle *Sense-Making* (strategic judgment) and *Accountability* (ownership and empathy).
2. From "Hierarchy of Authority" to "Hierarchy of Judgment": Decisions are no longer routed upward based on rank, but outward based on complexity (using the Cynefin framework).
3. From "Trust in Persons" to "Trust in Systems": Psychological safety relies on the transparency of the human-AI system (deliberation records, provenance chains) rather than just interpersonal dynamics.
4. From "Memory in Heads" to "Memory as Infrastructure": The "forgetting problem" of AI is solved by externalizing organizational memory into persistent, queryable state checkpoints.
5. From "Culture as Emergent" to "Culture as Designed": With AI handling routing, the casual interactions that build culture disappear. Culture must be intentionally engineered through Player-Coach mentorship and health monitoring.
6. From "Accountability as Blame" to "Accountability as Architecture": Accountability is built into the system via confidence-based routing and escalation paths, rather than sought after a failure occurs.

Phase 1: Architectural Mapping (The Protocol Design)

Before deploying agents, map the distribution of the Management Trinity. Identify which *Routing* functions the AI will automate, and explicitly assign the orphaned *Sense-Making* and *Accountability* functions to human roles.

The Management Trinity Decomposition

Role Redistribution

Anti-Patterns to Avoid

• The Hollow Middle: Removing managers without redistributing their functions. Leads to culture strain, burnout, isolation.
• The AI Manager: Assigning Sense-Making or Accountability to an AI agent. Leads to trust erosion, accountability vacuum.
• The Shadow Hierarchy: Informal leaders emerge to fill the gap, without formal authority. Leads to political dysfunction.
• The Overloaded DRI: Assigning too many cross-cutting problems to a single DRI. Leads to bottlenecks.

Phase 2: Governance and Guardrails (The Accountability Architecture)

Establish the systemic trust mechanisms that allow agents to operate safely.

1. Provenance Chains

A provenance chain links every agent action back to a human authorization. Every agent action must include:

• action_id, timestamp, agent_id, action_type
• human_authorizer (role, name, explicit authorization scope, date)
• inputs_considered, output, confidence_score

2. Confidence-Based Routing

Agents must express uncertainty as a resource.

3. Deliberation Records

A deliberation record captures the full reasoning process behind an agent's decision.

JSON Schema (deliberation_record.json):

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "Deliberation Record",
  "type": "object",
  "required": ["record_id", "timestamp", "agent_id", "provenance", "context", "alternatives", "rationale", "assumptions", "confidence", "limitations", "outcome"],
  "properties": {
    "record_id": { "type": "string", "format": "uuid" },
    "timestamp": { "type": "string", "format": "date-time" },
    "agent_id": { "type": "string" },
    "provenance": {
      "type": "object",
      "required": ["human_authorizer", "authorization_scope"],
      "properties": {
        "human_authorizer": { "type": "string" },
        "authorization_scope": { "type": "string" },
        "authorization_date": { "type": "string", "format": "date-time" }
      }
    },
    "context": {
      "type": "object",
      "properties": {
        "situation_summary": { "type": "string" },
        "data_sources": { "type": "array" },
        "cynefin_classification": { "type": "string", "enum": ["clear", "complicated", "complex", "chaotic"] }
      }
    },
    "alternatives": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "option": { "type": "string" },
          "pros": { "type": "array" },
          "cons": { "type": "array" },
          "risk_level": { "type": "string", "enum": ["low", "medium", "high"] }
        }
      }
    },
    "rationale": { "type": "string" },
    "assumptions": { "type": "array", "items": { "type": "string" } },
    "confidence": {
      "type": "object",
      "properties": {
        "score": { "type": "number", "minimum": 0.0, "maximum": 1.0 },
        "factors": { "type": "array" },
        "routing_action": { "type": "string", "enum": ["auto-executed", "human-reviewed", "escalated"] }
      }
    },
    "limitations": { "type": "array", "items": { "type": "string" } },
    "outcome": {
      "type": "object",
      "properties": {
        "action_taken": { "type": "string" },
        "result": { "type": "string" },
        "post_mortem_required": { "type": "boolean" }
      }
    }
  }
}

Phase 3: Collaborative Execution (The Judgment Topology)

Structure the day-to-day interaction between humans and AI.

The Cynefin Decision Router

• Clear: Apply best practices and automate. Human spot-checks.
• Complicated: AI gathers data, models scenarios, presents options. Human makes final decision.
• Complex: AI probes environment, senses patterns. Human responds adaptively.
• Chaotic: Human acts immediately to stabilize. AI used post-hoc for analysis.

AI Synthesis Report Template

When routing a decision to a human (70-90% confidence), the AI must present this report:

# AI Synthesis Report
**Decision Context:** [Clear / Complicated / Complex / Chaotic]
**Prepared For:** [Human DRI or Player-Coach Name]
**Date:** [YYYY-MM-DD] | **Agent ID:** [Agent identifier] | **Confidence Score:** [0.0 - 1.0]

## 1. Situation Summary
[Concise summary of the current situation and the key decision.]

## 2. Data Sources Consulted
| Source | Type | Relevance | Recency |
| :--- | :--- | :--- | :--- |
| [Source 1] | [Internal/External] | [High/Medium/Low] | [Date] |

## 3. Options Analysis
### Option A: [Name]
- **Description:** [What this option entails]
- **Pros:** [Key advantages] | **Cons:** [Key disadvantages] | **Risk Level:** [Low / Medium / High]

## 4. Assumptions Made
1. [Assumption 1]

## 5. Limitations of This Analysis
[Explicitly state what this analysis CANNOT account for, e.g., organizational context, politics.]

## 6. Recommendation (if confidence > 70%)
[Provide recommendation or state why human judgment is required.]

**Note to Decision-Maker:** This report is a starting point for your judgment, not a substitute for it.

Persistent Ownership Protocol

AI agents are stateless. To maintain ownership across sessions:

1. Session Checkpointing: At the end of every session, the agent generates a "State of the Project" summary (status, open questions, key context, human DRI).
2. Context Retrieval: At the start of a new session, the agent queries persistent storage (Vector Store or Graph DB) for the most recent checkpoint and historical context.

Phase 4: Human-Centric Maintenance (The Culture Engine)

Protect the emotional and psychological health of the organization.

The Player-Coach Model

• Player (60-70%): Building, shipping, executing. Uses AI agents as tools.
• Coach (30-40%): Mentoring, developing, connecting. Reviews AI Observation Reports.
• Workflow: AI tracks objective metrics (never keystrokes or sentiment) and generates a neutral Observation Report. The Coach synthesizes this data with context. The Coach delivers the mentorship session, focusing on career growth and psychological safety.

Trust Calibration

When AI errors occur, teams experience "trust ambiguity."

• Active Oversight: Implement friction points where humans explicitly validate AI outputs before proceeding. Rotate oversight to prevent complacency.
• Post-Mortem Protocol: Focus on the accountability architecture, not individual blame. Ask: Was the provenance chain intact? Was the confidence score accurate? Was the deliberation record adequate?

Culture Strain Prevention

Monitor for early warning indicators of culture strain:

• Isolation: Pulse survey score drops below 3.5/5. (Intervention: Increase Player-Coach touchpoints).
• Burnout: Pulse survey score rises above 3.5/5. (Intervention: Redistribute DRI load).
• Collaboration: Cross-team communication declines >20%. (Intervention: Reconnect team goals).

Self-Dependent Feedback Loop

This skill incorporates a dark factory intent engineering feedback loop that operates independently to continuously refine the organizational operating system.

Trinity Orchestrator Script (trinity_orchestrator.py)

Run this script to aggregate telemetry data across all six dimensions and generate improvement recommendations.

#!/usr/bin/env python3
"""
Trinity Orchestrator: Self-Improving Feedback Loop for the Management Trinity
Monitors: Accountability, Sense-Making, Trust, Ownership, Mentorship, Culture.
"""
import json
from datetime import datetime
from enum import Enum
from dataclasses import dataclass, field

class HealthStatus(Enum):
    HEALTHY = "healthy"
    WARNING = "warning"
    CRITICAL = "critical"

@dataclass
class DimensionHealth:
    dimension: str
    status: HealthStatus
    score: float
    indicators: dict = field(default_factory=dict)
    recommendations: list = field(default_factory=list)

@dataclass
class TrinityReport:
    timestamp: str
    overall_status: HealthStatus
    dimensions: list = field(default_factory=list)
    paradigm_shift_alerts: list = field(default_factory=list)
    improvement_actions: list = field(default_factory=list)

class TrinityOrchestrator:
    THRESHOLDS = {
        "accountability": {"escalation_rate_max": 0.30, "deliberation_quality_min": 0.80},
        "sense_making": {"ai_human_alignment_min": 0.60},
        "trust": {"psych_safety_score_min": 3.5, "cognitive_offloading_max": 0.20},
        "ownership": {"context_retrieval_success_min": 0.85},
        "mentorship": {"coach_time_allocation_min": 0.25},
        "culture": {"isolation_score_max": 3.5, "burnout_score_max": 3.5}
    }

    def assess_dimension(self, dimension: str, metrics: dict) -> DimensionHealth:
        thresholds = self.THRESHOLDS.get(dimension, {})
        recommendations = []
        issues = 0
        total_checks = len(thresholds)
        
        for metric_name, threshold in thresholds.items():
            actual = metrics.get(metric_name)
            if actual is None: continue
            if "max" in metric_name and actual > threshold:
                issues += 1
                recommendations.append(f"{metric_name}: {actual:.2f} exceeds threshold {threshold:.2f}.")
            elif "min" in metric_name and actual < threshold:
                issues += 1
                recommendations.append(f"{metric_name}: {actual:.2f} below threshold {threshold:.2f}.")
                
        score = 1.0 - (issues / max(1, total_checks))
        status = HealthStatus.HEALTHY if score >= 0.8 else (HealthStatus.WARNING if score >= 0.5 else HealthStatus.CRITICAL)
        return DimensionHealth(dimension, status, score, metrics, recommendations)

    def detect_paradigm_shift_alerts(self, dimensions: list) -> list:
        alerts = []
        culture = next((d for d in dimensions if d.dimension == "culture"), None)
        mentorship = next((d for d in dimensions if d.dimension == "mentorship"), None)
        trust = next((d for d in dimensions if d.dimension == "trust"), None)
        accountability = next((d for d in dimensions if d.dimension == "accountability"), None)
        
        if culture and mentorship and culture.status == HealthStatus.CRITICAL and mentorship.status == HealthStatus.CRITICAL:
            alerts.append("PARADIGM ALERT: 'Hollow Middle' anti-pattern detected. Immediate role redesign required.")
        if trust and accountability and trust.status != HealthStatus.HEALTHY and accountability.status != HealthStatus.HEALTHY:
            alerts.append("PARADIGM ALERT: Trust Collapse. Review deliberation record transparency.")
            
        return alerts

    def generate_report(self, all_metrics: dict) -> TrinityReport:
        dimensions = [self.assess_dimension(dim, all_metrics.get(dim, {})) for dim in self.THRESHOLDS.keys()]
        alerts = self.detect_paradigm_shift_alerts(dimensions)
        statuses = [d.status for d in dimensions]
        overall = HealthStatus.CRITICAL if HealthStatus.CRITICAL in statuses else (HealthStatus.WARNING if HealthStatus.WARNING in statuses else HealthStatus.HEALTHY)
        actions = [f"[{d.dimension}] {r}" for d in dimensions for r in d.recommendations]
        
        return TrinityReport(datetime.utcnow().isoformat(), overall, dimensions, alerts, actions)

if __name__ == "__main__":
    import sys
    if len(sys.argv) > 1:
        with open(sys.argv[1], "r") as f:
            metrics = json.load(f)
        report = TrinityOrchestrator().generate_report(metrics)
        print(f"Overall Status: {report.overall_status.value.upper()}")
        for alert in report.paradigm_shift_alerts: print(f">> {alert}")
        for action in report.improvement_actions: print(f"- {action}")