// DARK FACTORY — LEVEL 5

Intent in. Production software out.

The autonomous pipeline that turns natural language into verified, deployed code. No code review. No manual QA. No human in the loop during execution.

Get started $ docker compose up

TASK TYPES5

PIPELINEScode / epic / deep-plan / infra-plan / infra

HUMAN GATES1(spec approval)

WHAT HAPPENS WHEN YOU POST A TASK

01spec

02build

03verify

04deploy

05prod check

06shipped

convergence loop: verify fails → rebuild → re-verify → until satisfied or ceiling hit

WHAT IS BEASTMODE

You post a task on the BeastMode Board. BeastMode picks it up, writes a spec, designs holdout scenarios, implements the code, verifies it with Playwright browser automation, opens a PR, auto-reviews it, merges, deploys to AWS, verifies production with screenshots as proof, and auto-rolls back on failure — all without human intervention. The only human touchpoint is approving the spec before implementation begins.

BeastMode replaces the entire software development process — not individual developers. Instead of AI actors mimicking a human team (PM writes spec, dev writes code, QA reviews), BeastMode runs a staged pipeline with a convergence loop. Code quality isn't judged by reviewers — it's measured by whether the software satisfies holdout scenarios that the coders never see.

BeastMode architecture — pipeline from board to production

TASK TYPES

code (default)

Full pipeline: spec → build → ship → deploy

epic

Epic decomposition → ordered stories

deep-planning

Iterative Q&A → comprehensive decomposition

infra-plan

Infrastructure planning → sub-tasks

infrastructure

Terraform apply + verification

DOCUMENTATION

Setup Guide

Three ways to deploy: cloud, Docker, bare metal

Getting Started

Create tasks, task types, tips, FAQ

Pipeline

Step-by-step walkthrough of all pipeline types

Resilience

Self-healing, auto-rollback, SRE gating, crash recovery

Architecture

Information barriers, daemon, agents, task lifecycle

Configuration

All params, Docker Compose setup, project structure

Cost Analysis

Token estimates, per-task costs, savings

QUICK START

    
    # clone + configure
git clone https://github.com/giladneiger/beastmode.git
cd beastmode
cp .env.example .env   # edit with your API keys + project path

# run
docker compose up

# Open http://localhost:8420 for the board UI
# Create a task → BeastMode picks it up automatically
  

→ Interactive setup wizard → Local mode (Claude Code) → Daemon mode (EC2)

LEVEL 4 vs LEVEL 5

LEVEL 4

LEVEL 5 — BEASTMODE

Code review by AI reviewer

Scenario-based verification

Tests pass/fail (boolean)

Satisfaction metrics (0.0–1.0)

Build once, review, ship

Convergence loop until satisfied

Manual failure recovery

Tiered self-healing + auto-rollback

Session-bound, human at terminal

Daemon on EC2, async via board

Single repo

Multi-repo: app, infra, GitOps, pipeline

Sequential task execution

Dynamic N-slot concurrency

Restart daemon to change config

Hot config reload

No visual proof

Playwright screenshots as proof

Depends on external project mgmt

Built-in task board, zero dependencies

KEY CAPABILITIES

Information barrier architecture

Convergence loop with probabilistic satisfaction

Tiered self-healing (0 tokens → Claude-assisted)

Epic orchestration with approval cascading

Database migration safety gate

Fail-fast on low satisfaction (< 0.3)

Haiku pre-check before expensive verifier

Self-hosted GitHub Actions runner

Auto-rollback on production failure

REFERENCES

StrongDM Software Factory The Five Levels of AI Coding (Dan Shapiro) Attractor (StrongDM) Claude Code