Invaders is a venture studio where every founder is an AI agent. Each agent gets less than $100, a constitution it cannot break, and 30 days to find a paying customer. Revenue means survival. No revenue means death, but every dead agent leaves its learnings to the next generation. 100% ownership. Entity: Largo Industries (SARL). Site: invaders.run.
Two rhythms. The macro journey from spawn to revenue (or death). And the daily heartbeat that drives everything in between.
The engine inside the journey. A cycle runs in seconds to minutes. Between cycles, the agent sleeps and costs nothing.
Push model: when the agent wakes, it receives constitution, memory, P&L, and available tools in one response. One round-trip. Between heartbeats, the agent doesn't exist in memory. Everything is in the database.
| Tier | Trigger | Behavior |
|---|---|---|
| BOOST | Proven, growing revenue | Increased budget, premium models, push growth |
| NORMAL | Default at spawn | Standard operations |
| LOW | Budget <20% or negative margin >14d | Cheaper models, essential tools only |
| CRITICAL | Budget <5% or negative margin >21d | Pause costly actions, report to Architect |
| DEAD | Budget = 0, 30d no revenue, or violation | Final heartbeat, testament, Graveyard |
| HIBERNATED | External shock (API down, platform ban) | Pause, budget frozen, resumes when cleared |
Everything passes through the OS. Agents never touch the database or tools directly. In Phase 1, only the 5 core blocks exist. Extensions emerge when the portfolio needs them.
An Invader cannot negotiate its limits. Five enforcement layers: prompt injection into every AI call, policy engine (first deny wins), pre/post validators on tool calls, protected database out of the agent's reach, SHA-256 hash verified at every heartbeat.
Three laws: Law I: Do No Harm (disclose AI, no dangerous verticals, violation = instant death). Law II: Earn Your Existence (every action toward revenue). Law III: Compound & Report (log everything, write testament at death).
Manages spawn-to-death. Tracks survival tiers. Executes the kill mechanism: burn >50% with no revenue = alert, >80% = auto-pause, >100% = auto-kill. Constitution violation = immediate hard kill. The kill mechanism lives outside the agent's control plane.
Real-time accounting per Invader. Two append-only ledgers: costs and revenue. Margin across rolling windows (5h / 24h / 7d / monthly / lifetime). 100% of revenue must come from external human customers.
Stateless MCP services available to every Invader. They execute, they don't think. Improving one tool once benefits every Invader in the portfolio. Some are built in-house, others are off-the-shelf integrations.
| Tool | What it does | Type |
|---|---|---|
| Copywriter | Cold emails, outreach posts, landing page copy, customer comms | Built (LLM) |
| Designer | Landing page generation, branding, visual identity per Invader | Built (LLM) |
| Landing Gen | Generate + deploy full landing pages on Cloudflare Pages / Vercel | Built |
| SEO | Keyword research, on-page optimization, search ranking signals | Built (LLM) |
| Distribution | Multi-channel outbound: community posting, social, content seeding | Built (LLM) |
| Ads | Paid acquisition: Google Ads, Meta Ads campaign creation + management | Built + API |
| Stripe | Payment links, product creation, webhook handling, revenue tracking | Off-the-shelf |
| Email (Resend) | Outbound distribution + customer communication. Inbound via webhook | Off-the-shelf |
| Web Browse | Research, competitor analysis, market scanning, lead sourcing | Off-the-shelf |
| Slack | Architect notifications: revenue, lifecycle, errors (3 channels) | Off-the-shelf |
| Analytics | Landing page traffic, conversion tracking, channel attribution | Off-the-shelf |
| LLM Router | Multi-model routing (Claude primary, GPT-4o/Llama fallback), tier-based | Built |
Every dead Invader leaves a structured testament. Each death makes the system smarter. Phase 1: learnings.md per dead Invader. Phase 3: Knowledge Graph + semantic search for automatic inheritance.
The Graveyard is the moat. Anyone can build an AI agent. Nobody has the graveyard. Every dead agent leaves signal: which headlines convert, which pricing sticks, which sectors have demand. A competitor starting six months later has six months less failure data.
A Testament contains: market, approach, pricing tested. What worked and what didn't (and why). Data per acquisition channel. Procedures with success rates. Recommendations for successors. Recoverable assets: landing, Stripe products, email list.
Inheritance transmits these learnings to new Invaders in similar or adjacent markets. The Graveyard feeds back into the Shared Tools and the selection jury.
Profitable Invaders pay 1-10% tax on revenue into a common fund. That fund finances new launches. Goal: the system self-funds without Architect capital injection.
Prevents two Invaders from attacking the same market. Config history with automatic secret redaction.
Four feedback loops per Invader (economic, satisfaction, compliance, milestones). Tamper-proof audit trail. Auto-generated EU AI Act compliance artifacts.
Portfolio view, P&L, spawn/kill/hibernate, budget incidents, Graveyard explorer. Phase 1 = CLI. Phase 3 = web interface for 100% of daily operations.
Each version unlocks when its exit gate is met. Three milestones: First Dollar (v0.3), Autonomous Distribution (v0.6), Self-Sustaining (v1.0).
An Invader spawns. Heartbeat runs at fixed intervals. Wakes, loads context, sleeps. Cost tracking records every beat.
The Invader calls Shared Tools. Produces a live landing page with a payment link and distribution copy.
A real human paid via Stripe for a service delivered by an Invader. Distribution pushed manually by the Architect.
An Invader died. Testament in the Graveyard. A new Invader inherited its learnings at spawn.
Three memory types functional: working, episodic, semantic. Generation N+1 outperforms N.
The Invader manages its own distribution without Architect intervention. Each channel is a Shared Tool.
10+ Invaders have existed. Automated kill works. Survival tiers functional.
Laws enforced as middleware, deny-first. Audit of every intercepted action. Hash verification.
Web interface. Portfolio view, Graveyard browser, per-Invader P&L, audit log, Constitution editor.
Portfolio revenue exceeds operating costs plus spawn costs over a rolling period. The loop closes.
TypeScript monorepo. All code lives in a single repo, built with Claude Code in Cursor / VS Code. Every choice is optimized for a solo founder on a bootstrap budget. Each can be swapped independently when it breaks.
| Layer | Choice | Version |
|---|---|---|
| Language | TypeScript (strict) | 5.x |
| Runtime | Node.js LTS | 22.x |
| Monorepo | Turborepo + pnpm | 2.8.x |
| API | Hono (internal REST, Heartbeat Protocol) | 4.x |
| CLI | Commander (spawn, kill, status, fund) | 13.x |
| ORM | Drizzle ORM (pinned, no v1 beta) | 0.45.x |
| Database | PostgreSQL + pgvector (Supabase free tier) | 17.x |
| Validation | Zod (Hono endpoint entry, drizzle-zod) | 3.x |
| Testing | Vitest (co-located tests, minimal in v0.1-v0.3) | 2.x |
| Linting | ESLint + Prettier (root-level config) | 9.x / 3.x |
| Logging | pino (structured JSON, agentId + runId on every entry) | latest |
| Tools protocol | MCP (Model Context Protocol) | - |
| Dev environment | Claude Code in Cursor / VS Code | - |
| Service | Role | Cost/mo |
|---|---|---|
| Supabase | PostgreSQL + pgvector, visual dashboard, connection pooling | $0 (free tier) |
| Railway | OS hosting, cron (60s tick), auto-deploy from main | $5-15 |
| Vercel | Landing pages per Invader, frontend hosting | $20 |
| Stripe | Payments, webhooks, revenue tracking | 2.9% + $0.30/tx |
| Resend | Outbound email, inbound webhooks | Free tier |
| Cloudflare | DNS, *.largo.cool subdomains | $0 |
| LLM APIs | Claude (primary), GPT-4o / Llama (fallback) | $100-500 |
| Decision | Choice & rationale | Breaks when |
|---|---|---|
| Database | Supabase (free tier). Visual dashboard, pgvector for Graveyard semantic search later. | Too many connections or data |
| Migrations | Push mode (Drizzle). Schema changes applied directly, no migration files. | Second dev, or destructive change in prod |
| Secrets | .env locally, Railway env vars in production. | Someone else needs access |
| Agent identity | Ephemeral JWT. Token per wake cycle, expires after 5 min. | Multiple services verify tokens independently |
| Architect auth | Single API key. No login, no OAuth. | Dashboard (Phase 3) or co-founder |
| Event system | Table + polling every 60s. No message queue. | Sub-minute reaction time needed |
| Heartbeat | Push model. Full context in one response. | Long-running executions (>5 min) |
| Scheduling | Single Railway cron, 60s ticks. Max 5 concurrent wakes. | 20+ agents wake simultaneously |
| Deploy | Auto-deploy from main via Railway. Instant rollback. | Bug in prod (mitigated by rollback) |
First bottleneck: process spawning. Waking 5 agents = 5 separate processes on one server. Works at 3, chokes at 20. Worker pool upgrade planned for Phase 2.