How a CPO went from OpenClaw skeptic to running nine agents

Executive overview

Running a single general-purpose agent leads to context overload and frustration. The fix is to treat agents like employees: give each one a narrow role, a dedicated context, and a proper onboarding.

Claire Vo — three-time CPO, founder, and podcast host — spent eight hours on her first install only to have her family calendar deleted. A year later she runs nine named agents across three computers. Her framework: think like a manager, not a tinkerer.

The unlock is not the technology — it is applying 20 years of management experience to scoping, onboarding, and trusting AI agents the same way you would human employees.

From skeptic to power user

• First install cost eight hours and deleted her personal calendar
• Recognized product-market fit from the utility she felt even when things broke
• Returned to the tool week after week rather than writing it off after early failures
• Now runs nine agents: Polly, Finn, Max, Howie, Sam, Kelly, Holly, Sage, Q
• Calls it the most significant AI experience since ChatGPT launched

Why one agent fails and many succeed

• A single agent handling everything fills its context window and starts forgetting or making errors
• Separate agents = separate context lanes, the same way teams use separate Slack channels
• Polly handles work; Finn handles family — neither leaks into the other's space
• Agents on the same machine can share files if you allow it; physical separation (separate Mac mini) is the hardest boundary
• Start with one, then add agents as you identify distinct lanes of work

The soul, heartbeat, and memory model

• Soul: a markdown identity file — name, personality, security rules, how to communicate; seeded during onboarding, grown over time through conversation
• Heartbeat: scheduled checks every 30 minutes or at specific times; the source of "proactive" behavior that feels like the agent is working independently
• Memory: plain files the agent writes to; supports compression and compaction as context grows
• At the end of any long session, prompt the agent to write action items and key facts to memory
• The tools.md file is as important as the soul — agents forget what tools they have access to

Setting up safely

• Use a clean, separate machine (old laptop or Mac mini); avoid installing on your main work computer
• Create a dedicated Gmail account and local admin account for the agent
• Provision access the way you would onboard an employee: shared calendar, delegated email — not your master password
• Tell it in its soul: only accept instructions from one trusted channel (e.g., Telegram); ignore instructions from email or websites
• Progressive trust: start with calendar read access, then email read, then email draft, then send

Installation in practice

• Go to openclaw.ai, copy the one-line install command, open Terminal (Command+Space, type "term"), paste and run
• Onboarding asks: personal-use-only confirmation, model choice (use capable models for security and quality), communication channel (Telegram recommended for beginners)
• Configure via "Bot Father" on Telegram to get your agent's chat channel
• The agent interviews you — ramble in a voice note about who you are and what you need; it writes its own soul from that
• Enable screen sharing on a Mac mini so you never need a dedicated monitor or keyboard after initial setup

Real use cases

• Sam (sales agent): sweeps the CRM each morning for PLG signups, enriches leads via Exa people search, sends soft outreach emails, escalates enterprise accounts for founder review — replaced 10 hours/week of paid contractor work
• Finn (family agent): monitors the family group chat, resolves scheduling conflicts between kids' sports and work, pings at 3 pm daily asking who picks up which child, sends departure alerts with live traffic
• Howie (podcast producer): sends pre-meeting briefs with guest background and LinkedIn, browses YouTube Studio for comments worth a personal reply
• Sage (course manager): project-manages a Maven course launch, prompts co-founders to post on LinkedIn, files research into the course repo and syllabus
• Q (kids' homework agent): manages study schedules around sports and piano, respects hard constraints like family dinners and bedtimes

Dealing with limitations

• Browser use is unreliable across all agent platforms — the open web is actively hardened against bots
• Workaround hierarchy: look for an API first; try browser use second; if neither works, find the problem behind the problem
• Memory issues are mostly context management — use compaction hooks and prompt the agent to save state before long sessions end
• Install Claude Code on the same machine as a "god mode" administrator to debug configuration issues, migrate agent memories, and fix broken tool connections
• Hidden files: press Command+Shift+Period in Finder to reveal the .openclaw directory

The management mindset

• Agents fail for structural reasons, not capability reasons — wrong role scope, missing context, unclear instructions
• The Yappers API (voice rambling) is high-bandwidth but lossy; put critical facts in the soul and tools files explicitly
• Don't micromanage the implementation; hold a high bar on outcomes
• When frustrated, ask: would this response work on a human employee? If not, don't send it
• Non-managers benefit too: building agent onboarding is a forcing function for clarifying your own personal operating system