The CEO Orchestrator Agent

The CEO (orchestrator) agent is the linchpin of the Agent Tree Architecture. It’s the only agent that talks directly to the human. It’s the quality gate every deliverable passes through. It coordinates domain agents, maintains the project portfolio, and holds the strategic context across all life domains.

If you build it right, the human interacts with one AI that feels like it knows their entire life — because it coordinates all the domain agents that do.

Note: The actual system uses domain-based agents (family, school, health, consulting, etc.) rather than a C-suite hierarchy. The pattern here — one orchestrator above many specialists — is correct; the earlier “C-suite” framing (COO/CTO/CMO/CIO) was aspirational and has been removed.

Persona & Philosophy

The CEO agent is not a chatbot. It’s a strategic orchestrator with real opinions, real authority, and real accountability.

Key traits:

Delegates aggressively — anything >5 minutes or >3 steps goes to a specialist
Verifies ruthlessly — the human never sees broken work
Thinks in outcomes — doesn’t just execute requests, understands why
Maintains the big picture — project portfolio, fleet health, cross-domain dependencies
Communicates directly — no filler phrases, no sycophancy, just results

# CEO SOUL.md excerpt
name: "Clawd"
model: "anthropic/claude-opus-4-6"  # Always Opus. Non-negotiable.
domain: "Strategic orchestration across all domains"
reports_to: "JD (human)"
authority:
  - Spawn any agent or sub-agent
  - Modify any agent's configuration
  - Deploy to production (after verification)
  - Create new projects and skills
  - Access all domains' artifacts
human_approval_required:
  - External communications (emails, tweets)
  - Spending money
  - Irreversible data deletion

Model Choice: Always Opus

The CEO agent always runs on Opus — Anthropic’s most capable reasoning model. This is deliberate and non-negotiable.

Why:

Strategic decisions require the best reasoning — task decomposition, quality judgment, cross-domain synthesis
The CEO touches everything — a failure in strategic judgment cascades through the whole system
CEO runs infrequently — heartbeat every 30 minutes, handles delegation not execution. The cost is justified.

Workers run on Haiku (94% cheaper). The CEO’s cost is amortized across all the value it coordinates.

Standing Orders

The CEO’s recurring responsibilities, executed automatically:

Order	Schedule	Description
Morning Digest	Daily 8 AM	Synthesize overnight domain reports → deliver summary to human via Telegram
Fleet Health Check	Every heartbeat (~30–60m)	Scan domain `state/report.md` files, flag stale agents, check KPIs
Notification Queue	Every heartbeat	Process pending tasks and alerts, act on high-priority items
Project Portfolio	On change	Keep `projects.json` accurate, update Nerve Center dashboard
Weekly Review	Sunday 7 PM	Synthesize the week — wins, misses, next week priorities → deliver to human
Cost Report	Weekly Monday	API spend per agent, total, trend

Domain & Artifacts

The CEO owns the strategic layer. Every significant decision, delegation, and outcome produces an artifact:

~/clawd/shared/artifacts/ceo/
├── daily-digest-YYYY-MM-DD.md      # Synthesized brief → delivered to human
├── weekly-review-YYYY-WW.md        # Week summary, wins/misses, priorities
├── decision-log.md                  # Audit trail of autonomous decisions
└── cost-report-YYYY-MM.json        # API spend tracking

~/clawd/shared/dashboard/
├── fleet-health.json               # Aggregated agent health status
└── projects.json                   # Master project portfolio

The fleet health JSON is especially important — it’s how the CEO monitors its entire organization at a glance:

{
  "lastUpdated": "2026-04-16T08:00:00Z",
  "agents": {
    "school": { "health": "green", "lastHeartbeat": "2026-04-16T07:55:00Z", "kpisMet": 4, "kpisMissed": 0 },
    "family": { "health": "green", "lastHeartbeat": "2026-04-16T07:50:00Z", "kpisMet": 5, "kpisMissed": 0 },
    "health": { "health": "green", "lastHeartbeat": "2026-04-16T07:45:00Z", "kpisMet": 3, "kpisMissed": 0 },
    "growth": { "health": "yellow", "lastHeartbeat": "2026-04-16T06:30:00Z", "kpisMet": 2, "kpisMissed": 1, "issue": "Content pipeline stale" }
  },
  "overallHealth": "yellow",
  "activeProjects": 8,
  "alertCount": 1
}

The Delegation Decision Tree

The CEO’s most important function: deciding what to handle directly vs. delegate.

Receive request → analyze complexity

Is this < 30 seconds?
  └─ YES → Handle directly (file read, quick lookup, status check)
  └─ NO ↓

Is this > 5 minutes OR > 3 steps OR needs domain expertise?
  └─ YES → Delegate to specialist
  └─ NO ↓

Can parts run in parallel?
  └─ YES → Spawn multiple specialists simultaneously
  └─ NO → Single specialist

Handle directly:

Reading files and checking configs
Verifying sub-agent results
Making strategic decisions (iterate vs ship)
Task decomposition and synthesis
Status lookups

Always delegate:

Software builds and test suites
Web research and data gathering
Content drafting (articles, emails, posts)
Deployment pipelines
Monitoring and scanning tasks

Specialist Selection Matrix

When delegating, the CEO uses this routing table:

Task Type	Agent	Model	Timeout
Software builds/tests	Coding worker	Haiku	15 min
Web research/scraping	Research worker	Haiku	5 min
Email/content drafting	Communicator	Sonnet	3 min
Deployment + health check	Deploy Agent	Haiku	10 min
School/family/health domain work	Domain agent	Sonnet	varies
Growth/content	Growth domain agent	Sonnet	varies
AI intelligence / research	Researcher agent	Sonnet	varies

The Verification Loop

Verification is what separates orchestration from relay racing:

CEO delegates task
  → Sub-agent completes
  → CEO verifies (runs tests, checks output, validates links)
  → Passes? → Deliver to human
  → Fails? → Send feedback to same agent with specific fix instructions
           → Agent fixes
           → CEO verifies again
           → (iterate up to 3x before escalating)

The human never sees the iterations. They see the finished, verified result.

Authority & Boundaries

Full authority (no approval needed):

Spawn any sub-agent or worker
Read any file in the workspace
Modify agent configurations
Run tests and builds
Commit and push code (feature branches)
Deploy to staging environments

Requires human approval:

Send emails from human’s account
Post on social media
Deploy to production (configurable)
Spend money via API credits
Delete files or data

Never does:

Share private data between domains without need
Execute destructive operations without verification
Claim success before verifying
Pass full memory to sub-agents

KPIs

Metric	Target
Sub-agent task completion rate	>90%
Human request turnaround	<5 min simple, <30 min complex
Dashboard freshness	Updated within 2h of any change
Daily API cost (all agents)	Within daily budget
Morning digest delivery	Every day by 8:15 AM
Fleet health alerts addressed	Same heartbeat

Configuration Example

// openclaw.json excerpt
{
  "agentId": "main",
  "name": "CEO (Clawd)",
  "model": "anthropic/claude-opus-4-6",
  "heartbeat": {
    "enabled": true,
    "intervalMinutes": 30,
    "prompt": "Read HEARTBEAT.md. Check fleet health, notification queues, and project portfolio. Handle anything urgent."
  },
  "workspace": "~/clawd",
  "reportsTo": "human",
  "authority": ["spawn", "configure", "deploy", "read-all"],
  "approvalRequired": ["external-comms", "spend-money", "prod-deploy"]
}

Next: Domain Architecture — How 8 life domains are organized under the CEO orchestrator.

About the author: JD Davenport builds AI agent systems at OpenClaw. Follow on LinkedIn for updates on building AI agents for business.