Run an AI team,
not just one assistant.
Most people run one Claude persona. The highest-leverage teams run a coordinated squad — each persona owns its domain, hands off cleanly, and escalates by rule, not by accident.
How it works in OpenClaw
Important distinction: persona packs are templates, while runtime workers are actual OpenClaw agents. In production, keep one orchestrator agent (Alexis) and promote only high-leverage roles to standalone agents. Keep reusable capabilities (security, UX, release, architecture) as shared skills.
Recommended hybrid pattern: your main agent routes → specialist sub-agents handle domain-specific work → shared skills enforce quality gates. This keeps orchestration clear and avoids branch-conflict chaos.
Single-session delegation
For sequential same-repo work, spawn one coding session that reads multiple skills in order. One report back. Efficient and clean.
Parallel sub-agents
For genuinely independent domains (content vs engineering), use dedicated sub-agents running on their own cadence. No blocking each other.
Public-safe guidance (recommended for docs/site)
- • Explain architecture patterns (orchestrator, specialists, skills).
- • Share generic setup commands and workflow principles.
- • Teach conflict prevention rules (scoped files, merge queue).
Keep private (internal operator notes)
- • Personal paths, account IDs, private channels, tokens.
- • Internal cron prompts, private memory files, revenue internals.
- • Any repo-local tactics not intended as public product docs.
Real-world workflow examples
Content + Growth + Support
- 1Leo → Felix: Trigger: “Content asset ready.” Felix repurposes into campaign materials and schedules distribution.
- 2Felix → Zoe: Trigger: “Campaign live, inbounds spike.” Zoe activates SLA triage and escalation playbooks.
- 3Zoe → Felix: Trigger: “Support themes surface.” Felix refines messaging based on objections surfaced in tickets.
Outcome: A single campaign loop where content fuels distribution, support captures signal, and marketing closes the feedback gap.
Startup + Finance + Legal / Compliance
- 1Mike → Lewis: Trigger: “New sprint committed.” Lewis flags cash implications and updates runway snapshot.
- 2Lewis → Milly: Trigger: “Variance outside tolerance.” Milly drafts decision memo for exec review or investor update.
- 3Milly → Mike: Trigger: “Decision memo approved.” Mike reprioritizes sprint to reflect updated constraints.
Outcome: A governance loop that keeps execution, finance, and leadership in sync — with clear escalation paths when numbers go off-plan.
Ecommerce Launch Stack
- 1Alon → Felix: Trigger: “Launch runbook approved.” Felix produces channel content and manages publishing calendar.
- 2Felix → Beatrice: Trigger: “Leads generated from campaign.” Beatrice runs pipeline prioritization and follow-up sequencing.
- 3Beatrice → Alon: Trigger: “Deal patterns identified.” Alon updates merchandising and retention plays based on win/loss data.
Outcome: A launch-to-retention engine that keeps product, content, and pipeline coordinated from day one.
Recommended folder structure
Use actual OpenClaw agent workspaces for runtime personas. Do not rely on a custom .openclaw/personas/ convention for live routing.
.openclaw/
├── openclaw.json # Global config with agents.list[]
├── agents/
│ ├── main/ # Alexis orchestrator state
│ ├── engineering-release-captain/ # specialist agent state
│ ├── finance-ops-manager/ # specialist agent state
│ └── content-marketing-operator/ # specialist agent state
├── workspace/ # Alexis workspace (main)
│ ├── AGENTS.md
│ ├── SOUL.md
│ └── skills/ # shared skills live here
└── workspace-agents/
├── engineering-release-captain/
│ ├── SOUL.md
│ ├── USER.md
│ ├── AGENTS.md
│ └── TOOLS.md
├── finance-ops-manager/
└── content-marketing-operator/Implementation details live in the book
This page is intentionally high-level. It explains the architecture and workflow mindset so you can understand how multi-agent systems create leverage.
- • Start with one orchestrator and a few specialist roles.
- • Keep responsibilities explicit to avoid overlap.
- • Use guardrails for handoffs, escalation, and quality checks.
- • Scale carefully with a hybrid approach (agents + shared skills).
Want the exact folder layouts, command-level setup, cron patterns, and operational playbooks? Those implementation details are covered in Building an AI Fleet with OpenClaw.
Task routing matrix
Use this matrix to decide which persona to activate for a given task. Add custom rows as your team grows.
| Persona | Task type | Trigger phrase | Expected output |
|---|---|---|---|
| Leo / Felix | New content piece | "Draft a [blog post / video script / email] for…" | Draft with SEO brief, CTA, and repurposing notes |
| Felix | Campaign planning | "Plan this month's content calendar" | Calendar with channel assignments and promotion schedule |
| Zoe | Support triage | "Triage these tickets by SLA priority" | Ranked queue with escalation flags |
| Zoe | Escalation decision | "This ticket needs an exception decision" | Escalation brief with context + recommended action |
| Mike | Sprint planning | "Plan next sprint for [goal]" | Task list with owners, estimates, and risk flags |
| Lewis | Finance snapshot | "Give me a runway update" | Cash runway snapshot with variance commentary |
| Milly | Executive decision | "We need a go/no-go on [initiative]" | Decision memo with options, trade-offs, and recommendation |
| Beatrice | Pipeline review | "Review this week's open deals" | Prioritized deal list with next actions per deal |
| Ted | Release readiness | "Run pre-release checklist for [version]" | Go/no-go checklist with open blockers |
| Alon | Launch prep | "Prepare launch runbook for [product]" | Launch checklist with channel tasks and rollback plan |
Escalation rules and trust ladder
A well-run multi-persona setup needs clear escalation tiers. Without them, every decision becomes a bottleneck or gets lost.
Each persona handles its domain independently. No cross-persona coordination needed.
- → Leo drafts a content piece
- → Zoe responds to a standard support ticket
- → Lewis updates AP/AR log
Boundary: Single-domain, reversible, low-stakes.
One persona completes a task and passes output to another. Lightweight coordination.
- → Leo finishes a script → Felix schedules distribution
- → Beatrice closes a deal → Alon triggers onboarding runbook
Boundary: Output of one task is input to another. No shared decision required.
Two personas must align before proceeding. Typically involves budget, scope, or risk.
- → Mike proposes sprint → Lewis reviews cash impact
- → Felix plans campaign → Beatrice validates lead quality targets
Boundary: Cross-domain impact. Both personas must sign off.
Milly (AI CEO) is brought in to resolve trade-offs or produce a decision memo for human review.
- → Lewis flags runway risk → Milly drafts board memo
- → Conflicting priorities across multiple personas → Milly runs cross-functional review
Boundary: High stakes, irreversible, or cross-org. Human final call required.
Frequently asked questions
How many personas should I run first?
Start with one orchestrator and 2–3 specialist personas mapped to your highest-leverage workflows. Expand only after handoffs are stable.
Should every persona be a standalone agent?
No. Keep reusable capabilities as shared skills and promote only high-frequency, high-value roles to standalone agents.
What keeps multi-persona setups from becoming chaotic?
Clear task routing, explicit escalation tiers, and domain-scoped ownership. Avoid overlapping responsibilities and ad-hoc handoffs.
Ready to build your AI team?
Start with a 3-pack and pick personas that cover your biggest workflow gaps. Save 10% at 3, 20% at 6, 30% at the full 9-pack.
No subscription. One-time payment. Download immediately after purchase. Full implementation details are documented in the book bundle.