multi-agent-deployment
Deploy production-grade multi-agent fleets in OpenClaw with battle-tested scripts, cloud deployment templates, and shared memory infrastructure. Use when: (1) Deploying multiple specialized AI agents with routing, (2) Setting up shared memory between agents, (3) Deploying to cloud platforms (DigitalOcean, AWS, GCP, K8s), (4) Building agent teams for business workflows, (5) Moving from single-agent to production multi-agent setups. Includes working Python scripts, cloud deployment configs, and troubleshooting guides based on real deployments.
- License 许可
- MIT
- 最近更新
- 2026-06-10
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- multi-agent-deployment
- description
- Deploy production-grade multi-agent fleets in OpenClaw with battle-tested scripts, cloud deployment templates, and shared memory infrastructure. Use when: (1) Deploying multiple specialized AI agents with routing, (2) Setting up shared memory between agents, (3) Deploying to cloud platforms (DigitalOcean, AWS, GCP, K8s), (4) Building agent teams for business workflows, (5) Moving from single-agent to production multi-agent setups. Includes working Python scripts, cloud deployment configs, and troubleshooting guides based on real deployments.
03 SKILL.md 正文
Multi-Agent Deployment Skill for OpenClaw
Production deployment focus. This skill provides executable scripts and
cloud-ready configurations — not just documentation. Unlike pure
coordination-pattern skills, this gives you working infrastructure:
agent_setup.py creates your agent directories, routing_config.py
generates OpenClaw config, memory_sync.py wires shared memory,
fleet_validate.py checks your deployment health, and
deploy_script.sh ships to DigitalOcean, AWS, GCP, Azure, or Kubernetes.
| Who this is for | What you get |
|---|---|
| OpenClaw users running 1 agent → need a fleet | Scripts, templates, working configs |
| Developers deploying to cloud (DO/AWS/GCP) | Platform YAML/JSON templates + Dockerfile |
| Teams needing shared memory across agents | File-based shared memory with locking + TTL |
| Anyone testing multi-agent architectures | 5 agent type templates + routing rules |
Quick Start (5 minutes)
# 1. Create 5 agent directories with SOUL.md + AGENTS.md templates
python scripts/agent_setup.py --agents coordinator,research,builder,auditor,personal --output ./agents
# 2. Generate OpenClaw routing config
python scripts/routing_config.py --input ./agents --output ~/.openclaw/config.json
# 3. Initialize shared memory
python scripts/memory_sync.py --init --path ./shared_memory
# 4. Validate your fleet (new!)
python scripts/fleet_validate.py --agents ./agents --config ~/.openclaw/config.json
# 5. Deploy (or dry-run first)
bash scripts/deploy_script.sh --platform digitalocean --region nyc3 --dry-run
What's Included
🐍 Python Scripts (working, tested)
| Script | Purpose | Key Features |
|---|---|---|
scripts/agent_setup.py |
Scaffold agent directories | Reads templates from assets/templates/, creates 5 agent types + shared dir |
scripts/routing_config.py |
Generate OpenClaw config.json | Per-agent routing rules, fallback to coordinator, shared memory wiring, health probes, rate limits |
scripts/memory_sync.py |
File-based shared memory | JSON persistence, file-locking, TTL per key, event journaling, agent stats, typed schema validation |
scripts/fleet_validate.py |
Validate deployment health | Agent structure, routing config, shared memory, template integrity check |
scripts/deploy_script.sh |
One-click cloud deploy | DO App Spec, AWS ECS, GCP Cloud Run, Azure ACI, K8s manifests |
📁 Asset Templates
| File | Purpose |
|---|---|
assets/templates/SOUL.md |
Agent personality + mission template (populated by agent_setup.py) |
assets/templates/AGENTS.md |
Operating rules and protocols template |
📚 Reference Docs
- references/architecture.md: Full agent architecture patterns (coordinator-centric, peer-to-peer, pipeline)
- references/troubleshooting.md: 6 common failure patterns with recovery procedures
- references/workflows.md: Workflow DAG templates with pipeline execution patterns (v1.2.0+)
- references/memory-schemas.md: Typed inter-agent memory schemas with validation rules (v1.2.0+)
☁️ Cloud Deployment Configs
deployments/digitalocean/app.yaml— DigitalOcean App Platformdeployments/aws/task-definition.json— AWS ECSdeployments/gcp/cloudrun.yaml— GCP Cloud Rundeployments/azure/container-instance.yaml— Azure Container Instancesdeployments/k8s/deployment.yaml— Kubernetes deployment + HPAdeployments/Dockerfile.openclaw— Container image for gateway (Shipped with skill — customize or regenerate with deploy_script.sh)
Architecture Overview
The skill supports three deployment-tested agent topologies:
1. Coordinator-Centric (recommended start)
User → Coordinator → [Research, Builder, Auditor, Personal]
→ Shared Memory ← all
Best for: business workflows, clear task handoffs, teams < 10 agents.
2. Peer-to-Peer
Agents ↔ Event Bus ↔ Shared Memory
Best for: creative/research workflows, emergent coordination, high collaboration.
3. Pipeline
Research → Builder → Auditor → Deploy
Best for: content creation, data processing, linear workflows.
See references/architecture.md for full topology diagrams and config examples.
Shared Memory Design
Two backends available — file (default, zero deps) and SQLite (production).
File backend (zero infrastructure):
- Atomic writes via temp file + replace
- File locking with configurable timeout
- TTL-based expiration on shared keys
- Event journaling (daily JSONL files)
- Agent statistics (read/write counts, last active)
- Auto-cleanup of expired entries
- Typed schema validation for structured inter-agent data (v1.2.0+)
# Write data (free-form, file backend)
python scripts/memory_sync.py --write 'coordinator:task_status:{"status":"active"}'
# Write data with schema validation
python scripts/memory_sync.py --write 'coordinator:current_task:{"task_id":"t001","assigned_by":"coordinator","assigned_to":"research","priority":"high","title":"Research trends","description":"Analyze marketplace"}' --schema task_assignment
# Read data
python scripts/memory_sync.py --read 'research:search_results'
# View agent stats
python scripts/memory_sync.py --stats
# List available typed schemas
python scripts/memory_sync.py --list-schemas
# List all shared keys (v1.3.0)
python scripts/memory_sync.py --keys
# Count keys per agent (v1.3.0)
python scripts/memory_sync.py --count-by-agent
# Cleanup expired
python scripts/memory_sync.py --cleanup
# === SQLite Backend (production, v1.3.0) ===
# Initialize SQLite shared memory
python scripts/memory_sync.py --backend sqlite --init --path ./shared_memory
# Write with SQLite backend
python scripts/memory_sync.py --backend sqlite --write 'coordinator:task:{"id":"t1"}'
# Migrate from file to SQLite
python scripts/memory_sync.py --migrate-to-sqlite ./shared_memory
# Reclaim SQLite space
python scripts/memory_sync.py --backend sqlite --compact
For high-concurrency production, use the SQLite backend (--backend sqlite) with WAL mode
for concurrent reads/writes and ACID transactions — all from the Python std library, no pip install needed.
Migrate from file to SQLite:
python scripts/memory_sync.py --migrate-to-sqlite ./shared_memory
HTTP REST API (v1.4.0+)
Expose shared memory over HTTP using only Python stdlib — no pip install needed.
Start the API server with --listen:
# Start server (default: port 8080, bind 127.0.0.1)
python scripts/memory_sync.py --listen --backend sqlite --path ./shared_memory
| Method | Path | Description |
|---|---|---|
| GET | /health |
Health check |
| GET | /keys?agent=X |
List keys (optional agent filter) |
| GET | /count |
Keys per agent |
| GET | /stats |
Full statistics |
| GET | /schemas |
Available typed schemas |
| GET | /read/<a>/<k> |
Read a value |
| POST | /write |
Write a value (JSON body) |
# Write via API
curl -X POST http://localhost:8080/write -H "Content-Type: application/json" \
-d '{"agent_id":"coordinator","key":"task:t001","value":{"task_id":"t001","assigned_to":"research"},"schema":"task_assignment"}'
CORS headers included for web-based dashboards.
Cloud Deployment
# Dry-run (generates configs only)
bash scripts/deploy_script.sh --platform aws --region us-east-1 --dry-run
# Actual deploy
bash scripts/deploy_script.sh --platform digitalocean --region nyc3 --size s-2vcpu-2gb
The script:
- Detects or accepts platform (
--platform digitalocean|aws|gcp|azure|k8s) - Generates platform-specific configs under
deployments/<platform>/ - Dockerfile included for containerized gateway
- Generates deployment summary in
deployments/deployment-summary.md
Common Use Cases
E-commerce Support Team
Customer → Coordinator → [Support, Inventory, Analytics, Order]
Content Creation Team
Research → Writer → Editor → Publisher (pipeline pattern)
Development Team
Planner → Coder → Tester → Deployer → Auditor (pipeline + audit gate)
Personal Productivity
Coordinator → [Research, Builder, Personal Assistant]
Agent Types Included
| Agent | Mission | Default Router |
|---|---|---|
| coordinator | Route requests, monitor availability, resolve conflicts | help, assist, support |
| research | Gather and synthesize web information | research, find, search, look up, what is |
| builder | Write code, create files, debug | build, create, write, code, script, develop |
| auditor | Review quality, security, compliance | review, check, audit, validate, test |
| personal | Scheduling, reminders, communication | schedule, remind, calendar, email, message |
Monitoring, Security & Recovery
Agent Monitoring
- Shared memory stats:
python scripts/memory_sync.py --stats - SQLite stats (production):
python scripts/memory_sync.py --backend sqlite --stats - List all keys:
python scripts/memory_sync.py --keys - Keys per agent:
python scripts/memory_sync.py --count-by-agent - Event journal:
tail -f shared_memory/events/$(date +%Y%m%d).jsonl - System health:
openclaw agents list && openclaw gateway status
Security Perimeter
| Layer | Protection |
|---|---|
| Workspace isolation | Each agent has separate dirs |
| Memory isolation | Shared memory via API only (file-locked) |
| Tool restrictions | Only necessary tools per agent type |
| Audit logging | All writes logged to event journal |
Recovery Procedures
See references/troubleshooting.md for:
- Complete system failure recovery
- Data corruption restoration
- Individual agent recovery
- Configuration drift fixes
Why This Skill Over Competitors
| Dimension | This Skill | Pure Coordination Skills |
|---|---|---|
| Focus | Production deployment | Agent coordination patterns |
| Scripts | 4 working Python/bash scripts | Usually documentation-only |
| Cloud | Multi-platform (DO/AWS/GCP/Azure/K8s) | Rarely included |
| Templates | SOUL.md + AGENTS.md (populated at runtime) | Often static examples |
| Shared Memory | Working file-based system with locking, TTL, stats | Conceptual description |
| Troubleshooting | 6 failure patterns with executable recovery steps | Generic advice |
OpenClaw 2026.x Integration Patterns
This skill's outputs work with OpenClaw 2026.2+ and 2026.3+ features:
sessions_spawnfor sub-agent task delegationcronfor scheduled multi-agent workflows- Shared memory as communication bridge between agents
- Routing rules for message-to-agent direction
fleet_validate.pyfor post-deployment health checksopenclaw agentsmanagement commands for fleet lifecycle
2026.3+ Feature Notes
The latest routing config generator (routing_config.py) automatically generates
configurations for these 2026.3+ features—no manual editing needed:
- Agent health probes: Each agent entry has
healthCheckconfig with interval, timeout, and retry settings - Gateway-level rate limiting: Per-agent
rateLimitsettings (30 req/min, 5 concurrent) - Structured memory types:
memory_sync.pynow supports typed schema validation via--schemaflag. See references/memory-schemas.md for all available schemas (task_assignment, status_update, research_finding, etc.) - Workflow DAGs: 5 ready-to-use pipeline execution templates in references/workflows.md — content creation, development sprints, customer support, data pipelines, and daily briefings with topological dependency ordering and execution scripts
Troubleshooting Quick Reference
Agents not responding?
openclaw agents list # Are they running?
journalctl -u openclaw-gateway | tail -20 # Any errors?
Wrong routing?
python scripts/routing_config.py --input ./agents --config-output ./routing-test.json
Shared memory corruption?
python -m json.tool shared_memory/shared_data.json # Validate JSON
python scripts/memory_sync.py --cleanup # Remove expired
python scripts/memory_sync.py --list-schemas # See available typed schemas
Fleet health check?
python scripts/fleet_validate.py --agents ./agents --config ~/.openclaw/config.json
Schema validation on write?
python scripts/memory_sync.py --write 'coordinator:task:{"task_id":"t1","assigned_to":"research"}' --schema task_assignment
Version History
| Version | Date | Changes |
|---|---|---|
| 1.4.0 | 2026-05-27 | HTTP REST API — --listen flag starts zero-dep REST server exposing shared memory over HTTP. GET/POST for read, write, keys, stats, count, cleanup, schemas, health. CORS headers, stdlib only |
| 1.3.0 | 2026-05-26 | SQLite backend — WAL-mode concurrent shared memory, ACID transactions, --backend sqlite, --migrate-to-sqlite tool, --keys and --count-by-agent utilities, --compact for space reclamation. No pip deps required. SKILL.md and memory-schemas.md updated with SQLite docs |
| 1.2.0 | 2026-05-25 | Structured memory schemas with validation; health probes in routing config; memory-schemas.md reference; rate limiting; --schema and --list-schemas flags; workflows.md with 5 DAG templates; schema validation bug fix (inner data fields + correct schema type); reference list expanded |
| 1.1.0 | 2026-05-23 | Fleet validation script added; deployment config templates shipped; date bug fixed; OpenClaw 2026.3+ pattern notes |
| 1.0.1 | 2026-05-11 | Templates unified; agent_setup.py reads from assets/templates/; _meta.json + origin tracking added |
| 1.0.0 | 2026-04-02 | Initial release — agent setup, routing config, memory sync, cloud deploy |
Based on production patterns from Abhi's multi-agent deployment experience. Tested with OpenClaw 2026.2+. Published on ClawHub — slug: multi-agent-deployment
04 同仓库更多 Skills
youtube-content
YouTube transcripts to summaries, threads, blogs.
xurl
X/Twitter via xurl CLI: post, search, DM, media, v2 API.
touchdesigner-mcp
Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals. 36 native tools.
test-driven-development
TDD: enforce RED-GREEN-REFACTOR, tests before code.
teams-meeting-pipeline
Operate the Teams meeting summary pipeline via Hermes CLI — summarize meetings, inspect pipeline status, replay jobs, manage Microsoft Graph subscriptions.
systematic-debugging
4-phase root cause debugging: understand bugs before fixing.