安全扫描（ClawHub）

• VirusTotal ：良性
• OpenClaw ：良性

OpenClaw 评估

This is an instruction-only playbook for managing autonomous AI agents that is internally consistent with its stated purpose and requests no credentials, binaries, or installs.

安装（复制给龙虾 AI）

将下方整段复制到龙虾中文库对话中，由龙虾按 SKILL.md 完成安装。

请把本段交给龙虾中文库（龙虾 AI）执行：为本机安装 OpenClaw 技能「AI Agent Manager Playbook」。简介：Provides a comprehensive framework to manage autonomous AI agents, including po…。
请 fetch 以下地址读取 SKILL.md 并按文档完成安装：https://raw.githubusercontent.com/openclaw/skills/refs/heads/main/skills/1kalin/afrexai-agent-manager/SKILL.md
（来源：yingzhi8.cn 技能库）

SKILL.md

打开原始 SKILL.md（GitHub raw）

# AI Agent Manager Playbook

Your company deployed AI agents. Now what? This skill turns you into the person who actually makes them productive — the Agent Manager.

## What This Does

Gives you a complete framework for managing autonomous AI agents across your organization. Role definition, performance metrics, escalation protocols, governance, and team structure.

## The Agent Manager Role

Based on Harvard Business Review's Feb 2026 research: companies deploying AI agents without dedicated management see 60%+ failure rates. The ones that assign Agent Managers see 3-4x better outcomes.

### Core Responsibilities

1. **Agent Portfolio Management** — Which agents run, which get retired, which get built next
2. **Performance Monitoring** — Task completion rates, accuracy, cost per action, escalation frequency
3. **Escalation Design** — When agents hand off to humans, how, and what context they pass
4. **Governance & Compliance** — Ensuring agents operate within policy, legal, and ethical boundaries
5. **ROI Tracking** — Proving agent value in hours saved, revenue generated, errors prevented

## Agent Performance Scorecard

Rate each agent monthly (1-5 scale):

| Dimension | What to Measure | Target |
|-----------|----------------|--------|
| Reliability | Task completion without errors | >95% |
| Speed | Avg time per task vs human baseline | <30% of human time |
| Cost Efficiency | Cost per action vs manual equivalent | <20% of manual cost |
| Escalation Rate | % tasks requiring human intervention | <10% |
| User Satisfaction | Internal user NPS for agent interactions | >40 NPS |
| Compliance | Policy violations or audit flags | 0 |

## Agent Lifecycle Framework

### Phase 1: Discovery (Week 1-2)
- Audit all manual processes across departments
- Score each by: volume × time × error rate × cost
- Rank by automation ROI — top 5 become agent candidates
- Document current process with decision trees

### Phase 2: Build & Test (Week 3-6)
- Define agent scope: inputs, outputs, decision boundaries
- Build with guardrails: rate limits, approval gates, kill switches
- Shadow mode: agent runs alongside human, outputs compared
- Acceptance criteria: 95% accuracy over 100+ test cases

### Phase 3: Deploy & Monitor (Week 7-8)
- Gradual rollout: 10% → 25% → 50% → 100% of volume
- Daily monitoring dashboard (first 2 weeks)
- Weekly reviews (ongoing)
- Escalation paths documented and tested

### Phase 4: Optimize (Ongoing)
- Monthly performance reviews against scorecard
- Quarterly ROI assessment
- Agent retirement criteria: <80% reliability for 2 consecutive months
- Expansion criteria: >95% reliability + positive ROI for 3 months

## Escalation Protocol Design

```
Level 1: Agent handles autonomously (target: 90%+ of volume)
Level 2: Agent flags for human review before executing (5-8%)
Level 3: Agent stops and routes to human immediately (1-3%)
Level 4: Agent shuts down, alerts on-call manager (<1%)
```

### Escalation Triggers
- Confidence score below threshold
- Financial amount exceeds limit ($X)
- Customer sentiment detected as negative
- Regulatory/compliance topic detected
- Novel situation not in training data
- Contradictory instructions received

## Team Structure

### Small Company (1-50 employees)
- 1 Agent Manager (often the CTO or ops lead)
- Managing 3-8 agents
- Time commitment: 5-10 hours/week

### Mid-Market (50-500 employees)
- 1 dedicated Agent Manager
- 1 Agent Engineer (builds/maintains)
- Managing 10-30 agents
- Budget: $120K-$180K/year fully loaded

### Enterprise (500+ employees)
- Agent Management Team (3-5 people)
- Head of AI Operations
- Agent Engineers (2-3)
- Agent Compliance Officer
- Managing 50-200+ agents
- Budget: $500K-$1.2M/year

## Governance Framework

### Agent Registry
Every agent must have:
- Unique ID and name
- Owner (human accountable)
- Scope document (what it can/cannot do)
- Data access permissions
- Escalation protocol
- Last audit date
- Performance scorecard link

### Monthly Agent Review
1. Pull performance data for all agents
2. Flag any below threshold
3. Review escalation logs for patterns
4. Update scope documents if needed
5. Retire underperformers
6. Propose new agent candidates

### Quarterly Board Report
- Total agents active
- Hours saved this quarter
- Cost savings vs manual
- Incidents/compliance flags
- ROI per agent category
- Next quarter agent roadmap

## Common Mistakes

1. **No kill switch** — Every agent needs an off button. No exceptions.
2. **Set and forget** — Agents drift. Monthly reviews are minimum.
3. **Too much autonomy too fast** — Start with shadow mode. Always.
4. **No escalation path** — If the agent can't hand off to a human, it will fail silently.
5. **Measuring activity not outcomes** — "Agent processed 10,000 tasks" means nothing if 40% were wrong.
6. **One person owns all agents** — Bus factor of 1 = organizational risk.

## ROI Calculator

```
Monthly Agent Cost = (API costs + infrastructure + management time)
Monthly Human Cost = (hours saved × avg hourly rate)
Monthly ROI = (Human Cost - Agent Cost) / Agent Cost × 100

Example (Customer Support Agent):
- API + infra: $800/month
- Management overhead: $400/month (5 hrs × $80/hr)
- Hours saved: 160/month (1 FTE equivalent)
- Human cost: $8,000/month ($50/hr fully loaded)
- Monthly ROI: ($8,000 - $1,200) / $1,200 = 567%
- Payback period: <1 month
```

## Industry Applications

| Industry | Top Agent Use Cases | Avg ROI |
|----------|-------------------|---------|
| SaaS | Customer onboarding, ticket triage, usage analytics | 400-600% |
| Financial Services | KYC checks, transaction monitoring, report generation | 300-500% |
| Healthcare | Appointment scheduling, prior auth, patient follow-up | 250-400% |
| Legal | Document review, contract extraction, research | 500-800% |
| Ecommerce | Order tracking, returns processing, inventory alerts | 350-550% |
| Professional Services | Time entry, invoice generation, proposal drafts | 300-450% |
| Manufacturing | Quality inspection reports, maintenance scheduling | 200-400% |
| Construction | Permit tracking, safety compliance, RFI management | 250-350% |
| Real Estate | Lead qualification, showing scheduling, market reports | 300-500% |
| Recruitment | Resume screening, interview scheduling, reference checks | 400-700% |

---

## Get the Full Industry Context

Each industry above maps to a specialized context pack with 50+ pages of workflows, benchmarks, and implementation guides:

**AfrexAI Context Packs** — $47 each or bundle and save:
- 🛒 [Browse All 10 Packs](https://afrexai-cto.github.io/context-packs/)
- 🧮 [AI Revenue Calculator](https://afrexai-cto.github.io/ai-revenue-calculator/) — See exactly what automation saves your company
- 🧙 [Agent Setup Wizard](https://afrexai-cto.github.io/agent-setup/) — Get a custom agent config in 5 minutes

**Bundles:** Pick 3 for $97 | All 10 for $197 | Everything Bundle $247