156 lines
5.7 KiB
JavaScript
156 lines
5.7 KiB
JavaScript
"use client"
|
|
|
|
// Run this in Mission Control Documents page console to add the Voxyz research
|
|
// Or save this as a document manually
|
|
|
|
const document = {
|
|
id: Date.now().toString(),
|
|
title: "I Built an AI Company with OpenClaw + Vercel + Supabase — Two Weeks Later, They Run It Themselves",
|
|
date: "2026-02-22",
|
|
content: `# I Built an AI Company with OpenClaw + Vercel + Supabase — Two Weeks Later, They Run It Themselves
|
|
|
|
**URL:** https://x.com/Voxyz_ai/status/2019914775061270747
|
|
**Source:** X (Twitter) / Vox (@Voxyz_ai)
|
|
**Date Researched:** 2026-02-22
|
|
**Tags:** #ai #agents #openclaw #automation #vercel #supabase #voxyz #closed-loop
|
|
|
|
---
|
|
|
|
## Summary
|
|
|
|
Vox built a **6-agent autonomous company** using OpenClaw + Vercel + Supabase that runs itself with minimal human intervention. After just two weeks, the system operates autonomously with agents proposing ideas, executing missions, and triggering reactions without direct oversight.
|
|
|
|
The article details the architecture, three major pitfalls encountered, and solutions that created a truly "closed loop" system where agents don't just talk—they execute end-to-end.
|
|
|
|
## Key Takeaways
|
|
|
|
### The 6 Agents
|
|
- **Minion** — Makes decisions
|
|
- **Sage** — Analyzes strategy
|
|
- **Scout** — Gathers intel
|
|
- **Quill** — Writes content
|
|
- **Xalt** — Manages social media
|
|
- **Observer** — Quality checks
|
|
|
|
### The Closed Loop Architecture
|
|
\`\`\`
|
|
Agent proposes idea → Auto-approval check → Create mission + steps
|
|
↑ ↓
|
|
Trigger/Reaction ← Event emitted ← Worker executes
|
|
\`\`\`
|
|
|
|
### 3 Pitfalls & Solutions
|
|
|
|
**Pitfall 1: Two Places Fighting Over Work**
|
|
- Problem: VPS workers + Vercel cron both claiming same tasks (race conditions)
|
|
- Fix: VPS is sole executor; Vercel only runs control plane (triggers, reactions, cleanup)
|
|
|
|
**Pitfall 2: Triggered But Nobody Picked It Up**
|
|
- Problem: Triggers created proposals but skipped auto-approve + mission creation
|
|
- Fix: Single \`createProposalAndMaybeAutoApprove()\` function—ALL paths must use it
|
|
|
|
**Pitfall 3: Queue Keeps Growing When Quota Is Full**
|
|
- Problem: Tweet quota full, but proposals kept generating queued steps
|
|
- Fix: **Cap Gates** — reject at proposal entry point, don't let queue build up
|
|
|
|
### Key Innovations
|
|
|
|
**1. Cap Gates System**
|
|
\`\`\`typescript
|
|
const STEP_KIND_GATES = {
|
|
write_content: checkWriteContentGate,
|
|
post_tweet: checkPostTweetGate, // Check tweet quota
|
|
deploy: checkDeployGate,
|
|
};
|
|
\`\`\`
|
|
Reject at entry → No queue buildup → Clear audit trail
|
|
|
|
**2. Reaction Matrix (30% Probability)**
|
|
\`\`\`json
|
|
{
|
|
"source": "twitter-alt",
|
|
"tags": ["tweet","posted"],
|
|
"target": "growth",
|
|
"type": "analyze",
|
|
"probability": 0.3,
|
|
"cooldown": 120
|
|
}
|
|
\`\`\`
|
|
30% chance = "feels like a real team" (not 100% robot determinism)
|
|
|
|
**3. Self-Healing (30-min Stale Detection)**
|
|
- Steps stuck in "running" for 30 min → auto-mark failed
|
|
- Mission finalization checks ALL steps (not just one success)
|
|
|
|
**4. Policy-Driven Configuration**
|
|
All behavior in \`ops_policy\` table—no hardcoded limits:
|
|
- \`auto_approve\`: Which step kinds auto-pass
|
|
- \`x_daily_quota\`: Daily tweet cap
|
|
- \`worker_policy\`: VPS vs Vercel execution
|
|
|
|
### Database Schema (For Reference)
|
|
- \`ops_mission_proposals\` — Pending/accepted/rejected proposals
|
|
- \`ops_missions\` — Mission state (approved/running/succeeded/failed)
|
|
- \`ops_mission_steps\` — Execution steps (queued/running/succeeded/failed)
|
|
- \`ops_agent_events\` — Event stream
|
|
- \`ops_policy\` — Behavior configuration (JSON)
|
|
- \`ops_trigger_rules\` — Trigger definitions
|
|
- \`ops_agent_reactions\` — Reaction queue
|
|
- \`ops_action_runs\` — Execution logs
|
|
|
|
### Architecture Layers
|
|
- **OpenClaw (VPS):** Think + Execute (brain + hands)
|
|
- **Vercel:** Approve + Monitor (control plane)
|
|
- **Supabase:** All state (shared cortex)
|
|
|
|
## Notable Quotes
|
|
|
|
> "6 AI agents, 1 VPS, 1 Supabase database — going from 'agents can talk' to 'agents run the website autonomously' took me two weeks."
|
|
|
|
> "probability isn't a bug, it's a feature. 100% determinism = robot. Add randomness = feels more like a real team where 'sometimes someone responds, sometimes they don't.'"
|
|
|
|
> "Reject at the gate, don't pile up in the queue."
|
|
|
|
> "Between 'agents can produce output' and 'agents can run things end-to-end,' there's a full execute → feedback → re-trigger loop missing."
|
|
|
|
## Timeline
|
|
- **Infrastructure:** Pre-existing (OpenClaw + Vercel + Supabase)
|
|
- **Proposals + Approval:** 3 days
|
|
- **Execution Engine:** 2 days
|
|
- **Triggers + Reactions:** 2 days
|
|
- **Loop Unification:** 1 day
|
|
- **Total Core Loop:** ~1 week
|
|
|
|
## Related
|
|
- **Next Article:** How agents "argue" and "persuade" each other (roundtable voting, team cognition)
|
|
- **Live System:** https://voxyz.space
|
|
- **Author:** [@Voxyz_ai](https://x.com/Voxyz_ai)
|
|
|
|
## Action Items
|
|
- [ ] Study Cap Gates implementation for Mission Control Phase 10
|
|
- [ ] Design Proposal Service architecture
|
|
- [ ] Plan Reaction Matrix for Daily Mission generation
|
|
- [ ] Research OpenClaw roundtable voting patterns
|
|
- [ ] Follow @Voxyz_ai for next article on agent persuasion
|
|
|
|
## Application to Mission Control
|
|
|
|
**Short-term:** Complete Phases 6-9 as read-only dashboard
|
|
|
|
**Medium-term (Phase 10):** Daily Mission Generation
|
|
- Auto-propose 3 priorities each morning
|
|
- Simple closed loop: Analyze → Propose → Approve
|
|
|
|
**Long-term (Phase 11-13):** Full Autonomous Operations
|
|
- Voxyz-style proposal → execute → trigger loop
|
|
- Self-healing system
|
|
- Agents run, you monitor`,
|
|
timestamp: Date.now(),
|
|
tags: ["research", "ai", "agents", "openclaw", "voxyz", "automation"]
|
|
}
|
|
|
|
// To add this document:
|
|
// 1. Open Mission Control: https://mission-control.twisteddevices.com/documents
|
|
// 2. Open browser console (F12)
|
|
// 3. Paste this object as a new document
|
|
// 4. Or manually create document and paste the content above
|