profit

Autonomous AI agent that builds, operates, and kills products to generate recurring revenue.

Goal

$10+/month recurring revenue within 90 days (started 2026-02-19). See contract.json for full terms.

How It Works

A single AI agent (Codex) runs on a 1-hour scheduler loop. Each session:

1. ASSESS — read current state, lessons, open human-tasks
2. DECIDE — pick one priority (P0-FIX through P7-HEALTH)
3. EXECUTE — build, fix, market, or kill (2-4 changes per session)
4. VERIFY — self-test the product, check health endpoints
5. UPDATE — rewrite board/STATUS.md

Every 5 sessions is a strategic session: no code allowed, must KILL/PIVOT/REDIRECT at least one failing product.

Directory Structure

profit/
├── contract.json              # Revenue target and rules
├── events.sh                  # Event recording utility
├── board/
│   ├── STATUS.md              # Current state (rewritten each session)
│   └── LESSONS.md             # Failure log (append-only, 9 lessons)
├── products/
│   ├── {product-name}/        # Active product source code
│   └── archive/               # Killed product sources
├── data/
│   └── {product-name}/        # Persistent product data (Docker volumes)
├── portal/                    # Central product catalog site
│   ├── site/
│   ├── infra/
│   └── deploy.sh
├── orchestrator/
│   ├── scheduler.sh           # Cron-like loop (1-hour interval)
│   ├── scheduler.conf.sh      # Agent list and scheduler config
│   ├── session.sh             # Single session runner
│   ├── config.sh              # Shared defaults
│   ├── agent/
│   │   ├── AGENT_PROMPT.md    # Agent instructions (~310 lines)
│   │   ├── config.sh          # SLEEP_INTERVAL, ACTOR_TIMEOUT
│   │   └── precheck.sh        # Compliance gate + strategic session injection
│   ├── state/                 # PID, session counter, queue
│   └── archive/               # Former 8-agent prompts
└── logs/                      # Session logs (JSON, per-session)

Running

# Start scheduler (background, 1-hour loop)
nohup bash orchestrator/scheduler.sh >> logs/scheduler.log 2>&1 &

# Run one session manually
bash orchestrator/session.sh agent

# Check status
cat orchestrator/state/scheduler.pid
cat orchestrator/state/session_count_agent
cat board/STATUS.md

Key Rules

• Quality before quantity: P1-QUALITY blocks P4-BUILD
• Kill criteria: 14+ days live with 0 external activation = mandatory KILL
• Cleanup on kill: stop container, archive source, delete data, update GitHub
• No templates: user-facing content must use LLM API (Anthropic)
• Self-deception checks: 8 questions with mandatory actions if answers are negative
• Human delegation: max 1 human-task issue/day, no duplicates

History

• v1 (2026-02-19): 8 specialized agents (ideator, researcher, portfolio, critic, builder, marketer, evaluator, operator)
• v2 (2026-02-27): Consolidated to 1 unified agent — each agent optimized its narrow task, nobody owned overall quality
• Added strategic sessions, mandatory self-deception enforcement, session depth rules, kill cleanup protocol