Agent Deployment Checklist

Production deployment framework for AI agent infrastructure on dedicated hardware (Mac Mini, Linux servers). Every deployment follows the same 5-layer stack, every time, no shortcuts.

The 5-Layer Deployment Stack

Every agent deployment is five layers applied in order. No layer is optional. Each layer has a binary pass/fail gate before moving to the next.

Layer 1: Base OS + OpenClaw Install (Scripted)

Goal: Clean machine with OpenClaw runtime ready.

Checklist:

• - [ ] Fresh OS install or verified clean state
• [ ] OS updates applied to latest stable
• [ ] Xcode Command Line Tools installed (macOS)
• [ ] Homebrew installed and updated (macOS)
• [ ] Node.js LTS installed via nvm
• [ ] Python 3.11+ installed
• [ ] Git configured with deploy key
• [ ] OpenClaw CLI installed and verified
• [ ] Claude Code installed and licensed
• [ ] Working directory created at INLINECODE0
• [ ] SSH key pair generated for this machine

Script template:

CODEBLOCK0

Gate: node --version returns LTS, python3 --version returns 3.11+, git status works in workspace directory.

Layer 2: IAM Config (White-Glove)

Goal: Identity, access, and API keys configured for this specific client/deployment.

This layer is always done manually — never scripted — because every client's access pattern is different.

Checklist:

• - [ ] API keys provisioned (Anthropic, OpenAI if needed)
• [ ] API keys stored in environment variables (never in files)
• [ ] .env file created with proper permissions (chmod 600)
• [ ] Client-specific service accounts created
• [ ] MCP server credentials configured
• [ ] GitHub/GitLab access tokens scoped to client repos only
• [ ] Email/calendar integrations authorized (OAuth tokens)
• [ ] QuickBooks / accounting integrations connected (if applicable)
• [ ] All credentials tested with a live API call
• [ ] Credential rotation schedule documented

Key principle: Client pays for their own API keys and licenses. We never share keys across clients.

CODEBLOCK1

Gate: Every configured API key returns a valid response. No 401s, no 403s.

Layer 3: Client-Specific Software (Varies)

Goal: Install and configure whatever tools this specific client needs.

This layer varies per deployment. Common patterns:

For accounting/bookkeeping clients:

• - [ ] QuickBooks MCP server configured (read-only by default)
• [ ] Financial reporting templates deployed
• [ ] Tax calendar crons scheduled

For marketing/content clients:

• - [ ] CMS integrations connected
• [ ] Social media API access configured
• [ ] Analytics dashboards linked

For development team clients:

• - [ ] CI/CD pipeline access configured
• [ ] Code review automation set up
• [ ] Deployment notification channels connected

For legal/compliance clients:

• - [ ] Document management system access
• [ ] Compliance calendar configured
• [ ] Audit trail logging enabled

Gate: Client-specific test suite passes. Each integration returns expected data.

Layer 4: Security Hardening (Every Deployment)

Goal: Lock down the machine to production security standards.

Checklist:

• - [ ] Firewall enabled and configured (only required ports open)
• [ ] SSH hardening applied:

• - [ ] Disk encryption enabled (FileVault on macOS, LUKS on Linux)
• [ ] Automatic security updates enabled
• [ ] Fail2ban or equivalent installed and configured
• [ ] Log rotation configured
• [ ] File integrity monitoring enabled
• [ ] .env and credential files have 600 permissions
• [ ] No credentials in git history (verified with git log --all -p | grep -i "api_key\|secret\|password")
• [ ] SOUL, IDENTITY, USER, AGENTS files marked as sacred (never leave the environment)
• [ ] Outbound network allowlist configured (only known API endpoints)

macOS firewall script:

CODEBLOCK2

Gate: Security audit script returns grade A or B. No grade C or below passes.

Layer 5: Onboarding — Day 1

Goal: Client can interact with their agent and sees value immediately.

Checklist:

• - [ ] 5-file memory system scaffolded (see below)
• [ ] Starter crons installed and verified
• [ ] Health check running and reporting
• [ ] Client walkthrough completed (30-min live session)
• [ ] Client can ask agent a question and get a response
• [ ] First real task completed with client watching
• [ ] Emergency contact and escalation path documented
• [ ] Client has link to support channel
• [ ] Week-1 check-in scheduled

The Day-1 demo: Always do one real task live. Not a demo. Not a rehearsed script. Pick something from their actual workflow and do it. This is how you build trust.

Gate: Client has independently asked the agent a question and received a useful answer without help.

Pre-Scaffolded 5-File Memory System

Every deployment starts with the same five files. They are empty templates, not boilerplate — the agent fills them in during operation.

SOUL.md Template

CODEBLOCK3

IDENTITY.md Template

CODEBLOCK4

USER.md Template

CODEBLOCK5

AGENTS.md Template

CODEBLOCK6

MEMORY.md Template

CODEBLOCK7

Starter Cron Templates

Every deployment gets these three crons minimum.

Health Check (Every 4 Hours)

CODEBLOCK8

CODEBLOCK9

Memory Maintenance (Daily at 2 AM)

CODEBLOCK10

CODEBLOCK11

Backup (Daily at 3 AM)

CODEBLOCK12

CODEBLOCK13

Hardware Requirements

Minimum (Single-Agent Deployment)

Recommended (Multi-Agent Deployment)

Network Configuration

CODEBLOCK14

Post-Deployment Monitoring

Week 1: Daily Check-ins

• - Is the agent responding correctly?
• Are crons running on schedule?
• Any errors in logs?
• Client satisfaction?

Weeks 2-4: Twice-Weekly

• - Memory system growing appropriately?
• Performance stable?
• Any new integration needs?

Month 2+: Weekly

• - Health grade trend
• Backup verification
• Security update status
• Client feature requests

Deployment Anti-Patterns

Don't do these:

• - Sharing API keys across clients. Every client pays for their own. No exceptions.
• Skipping Layer 4. "It's just a test deployment" is how breaches start.
• Copying another client's SOUL.md. Every deployment gets a fresh identity calibrated to the client.
• Enabling write access on day 1. Start read-only. Earn write access through demonstrated reliability.
• Deploying without a health check cron. If you can't monitor it, don't deploy it.
• Promising specific features before Layer 3. Scope the deployment, then promise.