How It Works

Complete reference for the PowerLoop autonomous system

What Is PowerLoop?

PowerLoop is an Autonomous Product Studio — a system of AI agents that run on schedules to build, measure, and optimize a portfolio of products.

It uses the Karpathy autoresearch loop (measure → evaluate → modify → verify → learn → repeat) to continuously improve revenue across all projects. Each cycle, one variable is changed, the result is measured against a baseline, and the system decides whether to keep, discard, or modify the change.

The agents run autonomously via Claude Code scheduled tasks. They generate content, scan for leads, collect metrics, evaluate experiments, and compound learning across sessions through a shared Supabase database.

Architecture

┌─────────────────────────────────────────────────────┐
│              POWERLOOP DASHBOARD                      │
│  Portfolio · Agents · Experiments · Revenue · Tokens │
│  (What you're looking at right now)                  │
└──────────────────────┬──────────────────────────────┘
                       │ reads from
┌──────────────────────▼──────────────────────────────┐
│              SUPABASE DATABASE                        │
│  6 tables: experiments, revenue, strategy,           │
│  agent_runs, token_usage, projects                   │
│  Project: dvtyikhzieuqhcjglytf                       │
└──────────────────────┬──────────────────────────────┘
                       │ written to by
┌──────────────────────▼──────────────────────────────┐
│           AUTONOMOUS AGENTS                           │
│  (Claude Code scheduled tasks)                       │
│                                                       │
│  ORCHESTRATOR ──── evaluates + dispatches             │
│  METRICS COLLECTOR ── pulls data from everywhere     │
│  CONTENT ENGINE ──── generates + posts content       │
│  REVENUE SCOUT ───── finds opportunities             │
│  PRODUCT BUILDER ─── builds/deploys apps             │
│  PORTFOLIO MANAGER ─ maintains project health        │
└───────────────────────────────────────────────────────┘

Dashboard — This Next.js app. Reads from Supabase in real-time. Shows what the agents are doing.

Supabase — Persistent state. Every experiment, revenue event, strategy change, and agent run is recorded here. Agents communicate across sessions by reading and writing to these tables.

Agents — Claude Code sessions spawned by scheduled tasks. Each has a prompt defining its job. They run autonomously on cron schedules.

The Agents

Orchestrator

Weekdays 8:08 AM

The brain. Reads pending experiments, evaluates results against baselines, updates strategy parameters, and decides what to work on next. Dispatches other agents based on priorities.

Metrics Collector

Daily 6:00 AM

Pulls performance data from Stripe (revenue), Gmail (inbound leads), Netlify (traffic), and Supabase (user counts). Writes metrics to the revenue and experiments tables so the orchestrator has fresh data.

Content Engine

Tue/Wed/Thu 9:12 AM

Generates LinkedIn posts for LockedIn Labs using parameterized strategy (hook style, post length, tone). Each post is an experiment in the Karpathy loop — the strategy evolves based on engagement data.

Revenue Scout

Monday 7:06 AM

Scans Gmail for inbound consulting inquiries, searches the web for RFPs and opportunities, scores them by time-to-revenue, and creates Gmail drafts for high-priority leads. Never sends emails — always drafts for human review.

Product Builder

On-demand (dispatched by orchestrator)

When the orchestrator identifies something to build — a new feature, a pricing change, a micro-SaaS — this agent scaffolds, codes, and deploys it. Can create Next.js apps, Supabase schemas, and Stripe integrations.

Portfolio Manager

On-demand (dispatched by orchestrator)

Scans all projects for stale repos, broken builds, missing features, and growth opportunities. Feeds priorities back to the orchestrator.

The Karpathy Loop

1. MEASURE ──── Pull metrics for pending experiments
       │
2. EVALUATE ─── Score against baseline
       │         improved (≥10% better) → KEEP
       │         regressed (≥10% worse) → DISCARD
       │         neutral → MODIFY (try different variable)
       │
3. DECIDE ───── Update strategy parameters
       │         Commit winners, revert losers
       │
4. LEARN ────── Write findings to Supabase + memory files
       │         Compound knowledge across sessions
       │
5. PLAN NEXT ── Choose ONE variable to change
       │         Never repeat failed combinations
       │         Never change two things at once
       │
6. EXECUTE ──── Run the next experiment
       │         Log as "pending" for next cycle
       │
       └──────── REPEAT

One variable at a time. If you change the hook style AND the post length simultaneously, you can't tell which one caused the improvement. The loop enforces single-variable changes.

Never repeat failures. Every parameter combination that regressed is logged. The system checks history before proposing a new experiment to avoid repeating mistakes.

Compounding learning. Findings are written to both Supabase (for agent access) and project memory files (for session context). Effective workflows get captured as reusable SKILL.md files — the highest form of compounding.

Database Tables

superpower_experimentsEvery A/B test, content variant, pricing change, outreach attempt. Tracks hypothesis, variable changed, baseline vs result metrics, outcome, decision, and learnings.

superpower_revenueEvery dollar tracked from any source. Stripe payments, consulting leads, membership signups. Includes amount, source, type, and whether it's recurring.

superpower_strategyCurrent optimized parameters per domain (linkedin_content, moneyloop_saas, consulting_pipeline, outreach). Stores current values, full history, and target metrics.

superpower_agent_runsEvery autonomous agent execution. Status, actions taken, experiments created/evaluated, learnings, timing.

superpower_token_usageToken consumption per agent run. Input/output tokens, tool uses, duration, estimated cost in cents.

superpower_projectsMaster portfolio table. Every project with status, health, progress, revenue model, tech stack, links, and current MRR.

Revenue Streams

Project	Model	Status
MoneyLoop	SaaS — $39-79/mo subscriptions	Stripe live, awaiting first subscriber
LockedIn Labs	Consulting — $250-400/hr, $5K-25K packages	Service packages page live, content engine activating
HealthNext.ai	Enterprise — $100K-1M+ contracts	Site deployed, GTM in progress
KaleidoXDR	Endpoint licensing — $2-5/endpoint/mo	Product in development
Baraka Labs	Memberships — $9-199/mo	Membership page live with Stripe
Pixelizer	SaaS — $29-99/mo	Waitlist page with early-access pricing
Marketlink	Enterprise licensing via Wipro HPS	95% production-ready, ARC-AMPE compliance June 2026

Token Economics

The autonomous agents run on Claude Opus 4.6:

Input tokens$15 / million

Output tokens$75 / million

Context window1M tokens

Each scheduled agent run typically consumes 30K-100K tokens depending on complexity. The system tracks every run's token consumption and estimates cost, then compares total token spend against total revenue generated to calculate ROI.

The goal: every dollar spent on tokens should generate more than a dollar in revenue. The Tokens page shows this ratio in real-time.