Claire Vo's Multi-Agent OpenClaw Setup: What Actually Works (and What Deleted Her Calendar)

What Claire Vo Is Actually Doing

Claire Vo — host of *How I AI* and founder of ChatPRD — describes herself as one of OpenClaw's loudest early skeptics. On day one, she says, it deleted her family calendar. That's a useful detail to sit with before the rest of the story, because it frames what agent automation really is right now: powerful, occasionally destructive, and worth managing carefully.

Today she runs nine agents across three Mac Minis, each on a dedicated local account. The agents aren't a single do-everything assistant. They're split by function, more like Slack channels than one omniscient EA. A few of the named ones:

• Sam, a sales agent, does a daily CRM sweep, identifies decision-makers from new signups, and sends personalized outreach. Claire says it replaced a part-time salesperson she was paying roughly 10 hours a week.
• Finn, a home agent, pings her and her husband at 3pm every day asking who's picking up which kids, and flags scheduling conflicts (the oldest's basketball versus the middle kid's soccer) with a prompt to divide duties.
• Additional agents handle podcast prep, kids' homework help, and course project management.

That's the concrete picture. No benchmark claims, no funding numbers — just a working operator describing a setup she uses daily and comparing the feeling to "a ChatGPT moment."

The Ideas Worth Stealing

The setup details matter less than the operating principles, because those transfer regardless of which agent framework you pick.

Progressive trust. Claire onboards an agent the way you'd onboard a human executive assistant: first calendar access, then read-only email, then drafting, then sending. Given that her first-day experience involved a deleted calendar, this is less a philosophy and more a scar-tissue lesson. If you're wiring an agent into real accounts, don't hand it write and delete permissions on day one. Stage the access, watch the behavior, then expand. Isolation by design. A separate Gmail, a dedicated local account, and dedicated hardware (Mac Minis) keep the blast radius contained. When an agent misbehaves, it does so in a sandbox rather than in your primary inbox or your production CRM. This is the same instinct that leads teams building on n8n or Make to test flows against sandbox data before pointing them at live systems. Many agents, not one. The Slack-channels framing is the most useful mental model here. A single agent trying to be sales rep, family coordinator, and homework tutor at once accumulates context it doesn't need and blurs its permissions. Splitting by job keeps each agent's scope, tools, and access narrow — which is easier to reason about and safer to trust. Voice notes as the setup interface. Claire calls rambling into a voice memo "the yappers API" — the highest-bandwidth way to configure an agent. Instead of writing tidy specs, you talk through what you want in messy detail and let the model structure it. That's a genuinely different workflow from the trigger-and-action wiring you'd do in Zapier, and it hints at where agent configuration is heading. Claude Code as the repair tool. When things break, she reaches for Claude Code as a "brain surgeon" to fix and manage the OpenClaw setup. Worth noting for anyone assuming agents run themselves: they don't, and having a capable coding assistant on hand to debug the plumbing is part of the real cost.

Where This Fits — and the Honest Caveats

The headline claim from the conversation is that management skills matter more than technical skills. That rings true and also sets expectations. Running nine agents isn't a set-and-forget win; it's closer to managing a small team of overeager junior staff who occasionally do something catastrophic. You define the jobs, grant access carefully, review output, and correct course. If you're good at delegating to people, that skill transfers. If you're not, more agents will just create more mess faster.

Be clear-eyed about what's demonstrated versus what's aspirational. Sam replacing 10 hours of part-time sales work is one person's report, not a measured productivity study. The value here is the pattern — daily sweeps, decision-maker identification, staged outreach — which you could replicate on other stacks. Much of what Sam and Finn do (CRM checks, calendar conflict detection, scheduled prompts) is achievable with more conventional automation on n8n or Make plus an LLM step from OpenAI or Claude, often with more predictable behavior than a fully autonomous agent.

The interesting question isn't whether OpenClaw specifically is the answer. It's whether the agent-per-job, staged-trust, isolated-environment approach is the right way to build. On the evidence of a self-described skeptic now running it across her business and household, it's at least worth a careful experiment.

If you want help designing an agent setup with sensible permissions and a repair plan, browse the provider directory to find someone who can put it to work.