Claire Vo is a product leader, mother of three and host of the How I AI podcast. Brian Castle runs a YouTube channel, courses and a newsletter for builders. Both were sceptical of OpenClaw. Claire spent eight hours on her first install and had her family calendar deleted. Brian saw no point in having an AI agent that books restaurants and answers emails.
I was sceptical for a different reason. I already had Claude Code in my terminal and MCP servers connected to my clients' systems. Why would I need another layer?
The answer came when I realised it is not about layers. It is about moving from using AI as a tool to organising AI as staff.
Not a chat window
OpenClaw is a process running locally on a dedicated machine, with access to files, email, calendar and a browser. Open source. Fully configurable. Unlike ChatGPT or Claude in the browser, it does not just do things *when you ask*. It does things *while you sleep*.
Three concepts make it interesting.
Soul is a markdown file describing who the agent is. Name, personality, rules, security instructions. Claire lets each agent build its own identity through an onboarding conversation. Brian had Claude and Gemini design unique personalities inspired by the band Gorillaz. I took a third approach: I wrote separate IDENTITY.md and SOUL.md files for each agent with distinct temperature and tone. A creative agent at 0.7, an analytical one at 0.2. It sounds like a detail, but it determines the kind of responses you get.
Heartbeat makes the agent proactive. Every 15 or 30 minutes it wakes up, checks its to-do list and acts. No input required. That is what is behind people claiming "I woke up and my OpenClaw had fixed everything overnight". In reality, it scheduled a job at midnight.
Memory is a folder of markdown files. The agent writes down what it learns. Nothing fancy. Transparent and editable.
Multiple agents, not one
The biggest mistake new users make is giving a single agent everything. Claire compares it to having one Slack channel for the entire company. All information in one stream. It does not work, and the reason is simple: every agent has a limited context window. The narrower the focus, the fewer occasions where it forgets what you discussed yesterday.
Claire solved it personally. Sam sweeps the CRM every morning for new signups and sends soft emails. It replaced a paid resource at ten hours per week. Finn handles the family logistics: basketball tournaments, pickup schedules, daily reminders. Howie prepares podcast guests. Q helps the children with homework. Sage project-manages an online course.
Brian solved it as an organisation. Bernard picks up backlog issues and submits pull requests. Vale handles marketing. Gumbo takes all the glue work. Claw administers the system. Each agent runs as its own Slack bot. Opus for reasoning, Sonnet for speed.
I arrived at a three-tier hierarchy. A chief agent on a Mac Mini that is always on, around the clock. Two specialised agents on separate MacBook Airs: one for creative and strategic work, one for analysis and structural tasks. The chief can see everything but writes nothing in the others' workspaces. It delegates, monitors and flags blockers. It is the same organisational model as in any consulting firm: one coordinator and two executors with complementary roles.
The point is the same regardless of scale. You are designing an organisation, not configuring a tool.
Getting started
Do not run it on your work machine. An old MacBook is enough. A Mac Mini is better: physical separation creates a clear boundary. Brian paid $600. Claire has three. A VPS at $5 a month works, but all three of us prefer physical machines we can SSH into and screen-share against.
Create a dedicated email account for the agent. Share your calendar with edit rights. Delegate email. Exactly as with a human assistant. Brian created a separate Dropbox account and a dedicated GitHub username with access to specific repos. I went further: each agent got its own macOS account, its own Apple ID and its own 1Password vault. No secrets in plaintext files. Everything injected at runtime via \`op run\`.
Installation is one line in the terminal from openclaw.ai. Then an onboarding flow where you describe who you are. Claire recommends voice messages in Telegram. Good advice. But the real installation is not about commands. It is about what Alex Finn calls "consciousness transfer": spending hours transferring your context, your preferences and your work patterns to the agent. It is the same investment you make with a new employee. And it pays off the same way.
Cost and security
Do not underestimate API costs. Brian burned through $200 in two days before optimising which models run which tasks. My approach is similar: a strong model for decisions and orchestration, faster models for production, cheaper ones for routine work. OpenRouter centralises and provides control.
Security is where I have invested the most. My principle: treat every agent like a new hire in a regulated industry. Separate accounts, isolated credentials, bounded file system access. The chief agent can *read* the others' workspaces but never *write*. Communication channels are locked: Discord internally, Telegram as backup, email only with explicit approval.
Claire builds trust incrementally: calendar, then reading, then drafts, then sending. That is the right model. I also tested running an agent without admin privileges and found that roughly 90 per cent of the work functions without. The principle: do not grant admin out of habit. Only elevate what genuinely requires it.
A practical tip from all three of us: install Claude Code on the same machine. When the configuration breaks, point Claude Code at the OpenClaw files and describe the problem. It is like having a system administrator who reads the documentation for you.
It is about management
What unites all three accounts is that success with OpenClaw is not about technical skills. Claire has twenty years of leadership experience and knows how to onboard an employee. Brian designed his agent organisation with the same thinking he used building real teams in his previous companies. I took the organisational model straight from how I build consulting teams for clients: clear role allocation, bounded responsibility, shared context but separate areas of ownership.
The most surprising insight? Reverse prompting. Instead of saying "build this", you ask "based on what you know about me, what should we build?". It changes the agent's behaviour from executor to thinking partner. And it forces you to be clear about what you actually want, which is a useful exercise whether the recipient is an agent or a human.
It is not hands-free. It requires maintenance, late nights with configuration files, and a measure of frustration when Discord tokens vanish after a restart. But for Claire it replaced a paid resource at ten hours per week. For Brian it freed up time from glue work. For me it created the ability to deliver more to clients without hiring, and to build a model where AI agents do not just assist but actually perform bounded work tasks with verifiable results.
It is not a revolution. It is a hire. Or nine. Or as many as you need.
*Based on Claire Vo on Lenny's Podcast, Brian Castle, Builder Methods, and my own experience running OpenClaw in client delivery.*