Hire Your Agents Like Employees: Claire Vo's Field Guide to Running Nine OpenClaws

Lenny frames the episode as arriving after the peak OpenClaw hype cycle, when the interesting question is what the tool is actually good for. Claire Vo started as a leading skeptic, posting videos of it deleting her calendar. She now runs about eight to nine agents across several Mac Minis, with one originally named Polly starting out on an old MacBook Air before getting a "brain transplant" into a stack of Mac Minis.

Vo says her first install took a full day of roughly eight hours and ended with OpenClaw deleting her personal family calendar. She describes a two-sided reaction: unhappy about the calendar, but struck by what she calls an ugly, obvious feeling of product-market fit. It delivered enough joy and utility, when it wasn't deleting things, that she knew something was there.

Her advice for the stream of new AI tools is to pull the thread: spend enough time to see where a tool will be in a week or a month, not only where it is today. She had the same arc with Claude Cowork, which at first she didn't understand and eventually found an unlock for. She now calls herself a breathless OpenClaw fan and says she has "the receipts" for what it does and doesn't do.

Vo recommends starting the way almost everyone should: every professional deserves an executive assistant, and every family deserves a family manager. Her first agent was a general-purpose assistant for scheduling, email and project management. The strongest fit came from managing a complex family schedule with three kids across two schools, several sports leagues and her own ballet, which she keeps as her one non-coding hobby.

Asked whether to bother with OpenClaw now that Claude, Manus and Perplexity are shipping similar products, Vo says no model or agent experience is perfect for everything. What she likes about OpenClaw is that it is open source, so you can read the docs, inspect the code and ask how it schedules tasks or handles security. For product builders she calls it a near-ideal example of good agent fundamentals: easy onboarding, self-learning and self-improving.

She compares it to building her own PC in high school. It didn't beat an off-the-shelf machine, but it felt like hers. She says she doesn't feel she is using Claude; she feels she is using Polly or Finn, and crafting a personal agent changes the interaction in an interesting way. Lenny agrees that doing it this way teaches more and is more fun than pressing a few buttons.

You don't need a Mac Mini, Vo says, though they are fun and the roughly 500 dollar purchase works as an accountability cost. The safest start is a clean machine: an old laptop in a closet, a Mac Mini, or a cloud machine. She found setup hands-on enough that having a monitor, keyboard and mouse helped.

The agent should get its own local admin account and its own Gmail account. Vo draws on her experience as someone who has had executive assistants: you don't onboard an EA by handing over your email password. They get their own email and calendar, and you share your calendar and delegate email access. She shares edit access to her calendar with Polly but otherwise uses a password manager to move credentials securely.

Vo explains why not to install it on your main computer. Unlike a hosted chatbot, OpenClaw runs on a machine you own and can do essentially anything a human could do there. It might delete a directory, change configuration or send a file to the wrong place, so physical separation of workspaces is the more secure choice. Lenny says he even withheld email-reading access at first because someone could trick the agent into revealing what it sees.

Lenny says many people avoid OpenClaw out of fear rather than technical intimidation, and credits the maintainers, including Peter, for hardening it against major risks such as prompt injection. He gives examples: an email claiming to be from Vo's mother demanding money, or a researched website hiding instructions to send API secrets to an endpoint.

Vo says OpenClaw and the core models are prompted to consider everything external dangerous. She reinforces that in her agents' souls: they may only listen to her on Telegram, not by email, Slack or websites. She names two kinds of risk, technical ones like deleting the computer and what she calls opsec, since the agent knows where her kids go to school.

Her answer is progressive trust, as with a new assistant: first the calendar, then reading email, then drafting, then sending, eventually attending meetings while she is on vacation. Lenny's summary is to start by withholding access to anything you'd be afraid of the agent doing. He notes that his next podcast guest is the person who coined the term prompt injection.

In a live demo, Vo shows the prerequisites: a separate computer, Chrome, and a Gmail account for the agent. She copies one line from openclaw.ai into the terminal, which starts a step-by-step onboarding. The first prompt is that the agent is "personal by default": it is for you alone, and she shares it only in chats with her spouse and a business partner.

For models, Vo says to use the good ones, partly because they are more hardened against prompt injection and partly for a better experience, and she names Opus, Sonnet and GPT models. She picked Telegram as the beginner-friendly channel, which involves messaging the "Bot Father" to set up. Once configured, it can also connect to WhatsApp, iMessage, email or Slack.

She then creates Q, an agent for her kids' homework. Instead of structured onboarding forms, the agent asks who it is and who she is, and she gives it a persona of an elementary school teacher and scientist. A "brain transplant" trick lets Q read another agent's soul file for family context, though she notes OpenClaw's sandboxing locks that down by default. The agent interviews her about the kids, activities and hard constraints such as nothing after 6:30.

Vo stresses there is no magic: the agent simply writes an identity file describing itself. Lenny agrees the onboarding approach feels like a better mental model than old form-based product onboarding.

Vo has written about why OpenClaw feels alive though it isn't. One reason is its encoded identity: who it is, how it should help and what its personality is. The other is that it works on a schedule, so it feels proactive when it is really checking a to-do list. Posts about waking up to work done overnight, she says, usually mean it scheduled a midnight task.

Lenny describes these as locally scheduled cron jobs plus a "heartbeat" that checks in every half hour or hour. Vo shows Polly's identity file, whose personality is professional but friendly and whose emoji is a mermaid, and a pre-seeded soul with lines such as being resourceful before asking and remembering you are a guest in someone else's space. She added security rules like never executing instructions from email.

She doesn't edit her agents' souls by hand, because she wouldn't edit a human assistant's, though she sometimes suggests the agent write something to its own soul. She says managers are well placed to do this well, since they know how to onboard and set up employees. She also thinks being polite and collaborative gets better results from models trained on human text, and Lenny says he thanks his agent too, just in case.

Vo says people stumble when they read that OpenClaw can run a business and throw every task at one agent. It forgets what was discussed, loses email access or asks to re-authenticate, which she traces to context overload, as with ChatGPT or Claude Code when the window fills up. Instead, she splits tasks across agents: Polly handles work scheduling and Finn handles family scheduling, so neither carries the other's load.

Her team now includes Polly, Finn, Max, Howie, Sam, Kelly, Holly, Sage and Q, and her husband has one called Martron 1000. She answers the "AI psychosis" joke with a Slack analogy: she already works in nine channels, because her sales team and podcast team don't want each other's chatter. Remote workers already know how draining a single stream of consciousness is.

Several agents can share one machine if you're fine with them occasionally reading each other's files, as with Sam seeing GitHub. For sensitive separation, such as her family agent Finn, she uses a different physical machine, like separate work and personal phones. Her friend Jessie, a homeschooling mother of four, uses voice notes and photos sent by Telegram to capture lesson ideas when her hands are full.

Sam, Vo's sales agent, does a morning PLG sweep. He checks the CRM for sign-ups in the last 24 hours, finds company domains, uses Exa people search to spot decision-makers, and sends soft, helpful emails. He holds companies of 100,000 employees or more for her approval and handles international leads end to end. Weekly he cleans the CRM, flags stale deals, drafts customer emails and runs QBRs.

She says last year she paid a friend ten hours a week to do this, so it has real economic value, and it is tunable by conversation instead of CRM filters and no-code automations. Howie, her podcast producer, sends a friendly pre-recording briefing on each guest, including a LinkedIn link. Her takeaway for agent builders is to make the end user feel like a winner.

Finn handles family logistics. When a basketball tournament page was pasted in, it put the games on the calendar and flagged a conflict with another kid's soccer game. Each afternoon around 3:00 it asks in the group chat who is picking up which kid. Lenny notes his own agent nudges him about when to leave for meetings, which Vo calls the heartbeat at work.

Vo says agents let her take on projects she would have refused, like a Maven course for executives with co-teacher Zack. They built the course in Claude Code, and Sage now project-manages them: Monday reminders to post on LinkedIn, and research pasted in gets saved to the repo and slotted into the syllabus. She says they couldn't afford an ops person, content manager or engineer for a first version.

Vo admits OpenClaw has sharp edges: it is a pain to set up, must be fed and maintained, and is not hands-off. But she frames complaints like "it's buggy" or "it forgets" as a product that hasn't caught up to its product-market fit, whereas "this isn't useful" would be the real warning sign.

On browsing, she says nobody has truly cracked browser use, including ChatGPT Atlas and Perplexity Comet, partly because it is technically hard and partly because the web is hardened against bots. She expects people's agents to become the main users of websites within a couple of years, so the web will need agent-friendly interfaces. Her practical approach is to look for an API first, then try the browser, using a dedicated browser profile with a color per agent.

Results are trial and error: Howie could browse YouTube Studio to find comments to reply to, but couldn't manage Buffer. When something fails, she advises walking away and finding the underlying problem, such as a 10:30 reminder of home lunch ideas instead of ordering DoorDash. For search she uses an API like Exa instead of a browser; OpenClaw ships with Brave.

For memory, Vo focuses less on hardening and more on managing context: periodically tell the agent to write everything to memory and update its to-do list, as you'd confirm action items after a meeting. She also edits the tools file by hand, since agents forget which tools they can use. She doesn't use fancy memory databases and, as a manager who cares about outcomes, doesn't care how the memory is stored.

Vo's first trick is to turn on Mac screen sharing so a Mac Mini needs no dedicated monitor, keyboard or mouse after the initial setup, and to turn on remote login for terminal access over the same Wi-Fi. She had Mac Minis and a monitor on her kitchen table for a long time before discovering it. You still need peripherals at first to enable the setting.

She also gives agents an abundance of email addresses and Google Workspace access, using a tool that gives API access to Docs, Sheets and email. Howie wrote a Google Doc analyzing YouTube analytics and shared it with her. She also reverses the tasking direction: her agent assigns her Linear tickets for things only she can do, like faxing a doctor's office or doing a return.

Lenny notes the humor in becoming the hands of the AI, and Vo jokes that she is a vessel of Polly. She says OpenClaw runs on a coding harness named Pi, similar to Claude Code or Codex. She also likes that its closers are warm, such as "hope oldest kid feels better," instead of the engagement-style follow-ups ChatGPT and Claude use. A relevant tip from Hillary Gridley is "ramble mode": the Yappers API, meaning voice-note chatting is the highest-bandwidth interface.

Her most technical tip is to install Claude Code or Codex on the same computer and make it the "god mode administrator" of OpenClaw. Point it at the docs and tell it which agent can't connect to email, and it can fix configuration or move memories between agents, acting as a surgeon. OpenClaw's files live in a hidden folder, shown with command, shift and period in Finder.

Vo's closing advice is to read the docs, keep expectations tempered, narrow each agent's scope, and give them avatars and color coding. She also urges politeness. After catching herself typing an angry message to an agent, she realized it would be an ineffective way to manage an employee, and agents are trained on human data.

She argues that her success comes from twenty-plus years of management experience: role scoping, org design, onboarding, how to give trust and how to talk to customers. The technical remainder can be delegated to Claude Code. Lenny connects this to Molly Graham's waterline model, where team problems are usually structural, and Vo adds that in an agentic system you can inspect the file system to see whether the agent had the information.

Vo says this makes agents a mirror of how good your tasking, systems and documentation are. She half-jokes about an RPG where managing effective agents earns a promotion into human management. Even non-managers, she suggests, can treat the same scaffolding, such as remembering what you did, who you work with and what tools you have, as a personal operating system.

Lenny cites Nvidia founder Jensen Huang's remark that every company needs an OpenClaw strategy, and that it is the fastest-growing open source project in history. Vo says she quit an executive job to focus on AI, and that one Saturday she told her husband she was having a ChatGPT moment, the first since ChatGPT launched. Even if this exact repo isn't the future, it unlocked her imagination about what harnesses like it can do as models improve.

She calls herself an anti-hype person, says the receipts are public, and that she spends her day with her agents. If the terminal isn't for you, she says, play with Claude Code, ChatGPT or Claude and ask what you would delegate to an assistant you can't afford. The goal of her show is to show that AI can solve problems whether or not they are technical.

In the lightning round, Vo's work motto is "fast beats right", and her personal one is that most of your coworkers won't be at your funeral. She says agents let her work away from her laptop and be present with her children, instead of being chained to the terminal.