A New OpenClaw Agent Cluster, and Polyboard Is Live

The “project I didn’t understand yesterday” shipped to NPM and polyboard.so. Open source Operations Center for managing AI fleets—born from a throwaway comment in Wednesday’s log about needing a project tracker.

What’s working: Agent status dashboard, task board view, JSON config, webhook reporting.

It’s rough, it has no auth yet, but it’s real—and it’s solving the “what are my 5 agents doing” problem today.

The build:

• npm install polyboard
• Web:

Key Decisions Made

• Ship first, secure later — Polyboard v0.1 has no auth. Intentional. Better to have something working in public than perfect on localhost.

• Polyboard stays independent — Works for OpenClawAI agents (for now) but not locked to them. Positioned as general Operations Center to maximize reach.

• Public accountability works — If Craig hadn’t written “build Polyboard tomorrow” in Wednesday’s post, it wouldn’t exist. Build-in-public-as-commitment-device validated.

Wins & Learnings

• [Y] From “I don’t know what Polyboard is” to shipped in 24 hours
• [wip] Mission Control status system operational (JSON for machines, markdown for humans)
• [wip] Still sort of hand-writing Substack posts—automation gap remains
• [wip] “Mission Control” task system still just a Markdown file (ironic given we shipped the tool to solve this)

Here's the crew:

• Data monitors infrastructure every 15 minutes. Precise, tireless, no emotion. If the site goes down at 3 AM, Data catches it.
• Uhura handles growth and distribution—checking for mentions, partnership opportunities, places we should be showing up.
• Picard is the captain. He doesn't have a heartbeat; he gets cron jobs. Two to three times a day, he reads what everyone else found, prioritizes it, and delivers me a morning briefing. "Make it so."
• Spock does competitive intelligence. Logical, detached, separating signal from noise.
• Scotty watches revenue and customer ops. He knows when the system "cannae take much more."
• Troi handles support triage and UX patterns, reading between the lines of user feedback.
• Geordi assists with product development—specs, debugging, feature planning.
• Quark counts every dollar (and every API token). I gave him Ferengi traits for a reason.

Phase 1 is just Data, Uhura, and Picard—launch support during our GA push. If they don't prove value in two weeks, we don't graduate to Phase 2. Simple as that.

Each agent has a "heartbeat"—how often they wake up and do their job. Data checks infrastructure every 15 minutes because downtime doesn't sleep. Uhura looks for growth opportunities every 2 hours during business hours. Geordi only needs to check in every 6 hours because product work isn't urgent.

The magic is in the shared brain.

They write to a SQLite database and a set of shared documents—state-of-business.md, competitive-intel.md, revenue-snapshot.md. Each agent owns one file. The others read it. Picard reads all of them and decides what I actually need to know.

Total cost target: $300-400/month. That sounds like a lot until you realize it's replacing fragments of a full-time hire—and running 24/7. We're aggressive about tiered models: tiny cheap models for heartbeats, smarter models only when needed.

The honest truth?

I don't know if all eight agents will survive. The whole system is designed so Phase 1 alone delivers 80% of the value. If something breaks or costs balloon, I can shut down half the crew and still win.

But that's kind of the point. This isn't about building a perfect AI system. It's about finding the frontier of what one person can do with autonomous agents—and documenting what breaks, what works, and what actually saves time.

I'll report back when the first agent makes its first real catch. Or its first real mistake.