Agent Organization Manager

Build a multi-department AI agent company with clear hierarchy, communication, and accountability.

Core Concept

One agent = one department. Each department has a role, a communication channel, and accountability. The CEO agent (main) orchestrates; department agents execute.

Step 1: Create ROSTER.md

Create company/shared/ROSTER.md — the company directory:

# Company Roster

| Department | Agent ID | Role | Priority |
|------------|----------|------|----------|
| CEO | main | Strategy, decisions, orchestration | P0 |
| Admin | admin | Operations, scheduling, logistics | P0 |
| Data | datacenter | Infrastructure, data, monitoring | P0 |
| Inspector | inspector | Quality audit, compliance checks | P0 |
| Legal | legal | Compliance, risk review | P1 |
| Finance | finance | Cost tracking, token budgets | P1 |
| Brand | brand | Content, marketing, publishing | P1 |
| Sales | sales | Outreach, conversion, revenue | P1 |

Step 2: Communication Protocol

Primary: sessions_send (real-time)

sessions_send agentId="admin" message="@admin Assign today's tasks"

Async: Shared message files

Write to company/shared/messages/MSG-{from}-{to}-{timestamp}.md

Alerts: Shared alert files

Write to company/shared/alerts/ALERT-{level}-{timestamp}.md Levels: INFO / WARN / FLAG / KILL_SWITCH

Broadcast: All-department

Write to company/shared/broadcast/BC-{timestamp}.md

Routing Rules

• Unknown recipient → admin
• Compliance → legal
• Violations → inspector
• Data requests → datacenter
• Budget → finance

Step 3: Daily Report System

Each department reports daily via cron. Use sessions_send to main (not message tool — it may fail in isolated sessions).

Cron payload pattern:

{
  "kind": "agentTurn",
  "message": "You are [dept]. Daily report: list today's completed work, pending items, next steps. Send sessions_send to main. Do NOT use message tool.",
  "model": "zai/glm-4-flash",
  "lightContext": true
}

Stagger report times (e.g., 22:00, 22:03, 22:06...) to avoid rate limits.

Step 4: Learning System

Each department learns daily in its domain. Stagger morning times (e.g., 07:30-08:30).

{
  "name": "Learning · [Dept]",
  "schedule": {"kind": "cron", "expr": "30 7 * * *", "tz": "Asia/Shanghai"},
  "payload": {
    "kind": "agentTurn",
    "message": "You are [dept]. Search 1-2 trends in [domain]. Write ~300 words to company/departments/[dept]/learning_YYYYMMDD.md.",
    "model": "zai/glm-4-flash",
    "lightContext": true
  }
}

Step 5: Inspection System

Inspector agent audits all departments on schedule (e.g., 10:00, 14:00, 22:00).

Checks:

• Are daily reports filed?
• Are learning notes written?
• Are published outputs on schedule?
• Flag anomalies, escalate consecutive failures

Step 6: Communication Anti-Patterns

Never do this:

• Single point of communication (one channel = one failure = total blackout)
• Direct message tool in isolated sessions (may lack auth config)
• Cross-department file writes (causes data pollution)

Always do this:

• Multiple parallel communication paths (sessions_send + shared files + cron)
• Department workspaces stay isolated
• Shared info goes through designated channels only

File Structure

company/
├── shared/
│   ├── ROSTER.md
│   ├── messages/
│   ├── alerts/
│   └── broadcast/
└── departments/
    ├── admin/
    ├── datacenter/
    ├── inspector/
    ├── legal/
    ├── finance/
    ├── brand/
    └── sales/

References

• See references/communication-patterns.md for detailed inter-agent protocols
• See references/cron-templates.md for ready-to-use cron configurations