Zero Factory: A 24/7 AI Multi-Agent Software Factory

Zero Factory: Introduction to the 24/7 AI Multi-Agent Software Factory

Zero Factory is an AI multi-agent orchestration system released by hotcode-dev on GitHub on June 14, 2026 (original link: https://github.com/hotcode-dev/zerofactory). Built on Hermes Agent, it consists of six specialized agents forming a complete software factory, realizing an end-to-end automated pipeline from requirement research to code deployment. Core features include: 24/7 uninterrupted development, parallel execution capability, human review at key nodes, cost optimization strategies, etc.

This thread will introduce its architecture, workflow, optimization strategies, and practical significance in detail across separate floors.

Background and Core Team Architecture

What is Zero Factory?

Zero Factory is an AI multi-agent orchestration system built on Hermes Agent. Its core concept is 24/7 uninterrupted development—through frequent re-prioritization, rolling handoffs, and parallel execution, it keeps the software development process running continuously. Unlike traditional single-agent tools, it breaks down development links into independent roles, with each agent having a specialized skill set and personality configuration to maximize parallel efficiency and output quality.

Six Core Team Members

1. Orchestrator: Process supervisor, monitors the kanban, manages task decomposition, coordinates handoffs, and escalates issues to humans when blocked.
2. Researcher: Technical architect, explores APIs, researches optimal solutions before coding, and writes detailed technical specifications.
3. Builder: Senior software engineer, focuses on code writing, emphasizing speed, type safety, test coverage, and feature delivery.
4. Reviewer: Quality gatekeeper, reviews Pull Requests, identifies bugs, performance issues, and security vulnerabilities.
5. QA: Verifies edge cases, runs integration tests, and ensures code meets acceptance criteria.
6. Scribe: Maintains README, API documentation, and operation manuals, ensuring documents are updated in sync with code.

Workflow and Parallel Execution Mechanism

Workflow State Transitions

Based on the Hermes Kanban system, the state transitions are: Triage → Todo → Ready → In Progress → Blocked → Done

• Triage: Initial goal received, automatically split into subtasks
• Todo: Subtasks generated, waiting for human plan review
• Ready: Approved for execution, automatically assigned by the scheduler
• In Progress: Agent actively processing
• Blocked: PR created/waiting for human review
• Done: Task completed and archivable

Key Human Review Nodes

1. Plan Review: Before Todo→Ready, humans review and approve the task plan
2. Result Review: After the Reviewer passes, the task enters Blocked to wait for final human acceptance

Parallel Execution Mechanism

• Task Dependency Management: Parent-child task relationships; subtasks unlock only after the parent task is completed (using the --parent flag)
• Auto Scheduling: The kanban-dispatcher plugin assigns Ready tasks based on agent availability
• State Isolation: Agents have independent configurations and tool sets to avoid interference

Cost Optimization Strategies and Configuration Extension

Cost Optimization Strategies

Technology	Setting	Purpose
Tool Set Isolation	Per config file	Each agent only has the required tools
Context Compression	Enabled, threshold 0.5	Reduce context window usage
Prompt Caching	Cache TTL 5 minutes	Reuse system prompt tokens
Max Round Limit	60-120 per config file	Limit token consumption per task
Short Lifecycle Tasks	Task-based lifecycle	No resource consumption by idle agents

Configuration and Extension

• Multi-layer Configuration: Base configuration (profiles/common/config.yaml) + role-customized configuration (profiles/<profile>/config.custom.yaml)
• Custom Skills: Add skills in profiles/common/skills/, run make skills-link to link to all configurations
• Apply Changes: After modifying SOUL files (agent personality definitions) or configurations, run make merge-all to regenerate configurations

Communication Mechanisms and Deployment Requirements

Agent Communication Channels

1. Kanban: Task handoff, state tracking, dependency management
2. File System: Shared output files, configurations, documents
3. Gateway API: Real-time message delivery

Tech Stack and Deployment

• Environment Requirements: Linux machine (Raspberry Pi/x86_64), Python 3.11+, Bun 1.1+, Hermes Agent installed
• Deployment Steps: Clone the repository → Link Hermes configuration files → Merge configurations → Start the Gateway

Core Design Principles and Practical Significance

Eight Core Principles

1. 24/7 Development: Continuous iteration with frequent re-prioritization
2. Productivity and Automation: End-to-end automation of routine tasks
3. Quality and Reliability: Layered reviews ensure high-quality, maintainable, and secure software
4. Cost Efficiency: Optimize token usage; each agent only has necessary skills
5. Hybrid Review: AI-assisted review at each stage, human intervention for key decisions
6. Single Source of Truth: Standard location for shared information; link instead of copy
7. Minimalism: As small, simple, clean, and usable as possible
8. Living Documents: Documents updated in sync with code changes

Practical Significance

• Role Specialization: Division of labor allows agents to perform better in specific areas, improving collaboration efficiency
• Process Standardization: Kanban states and handoff mechanisms make implicit knowledge explicit, forming repeatable and optimizable processes
• Human-AI Collaboration: Human reviews at key nodes balance automation and human control
• Cost Controllability: Optimization strategies make continuous operation economically feasible
• Reference Value: Provides a reference architecture for teams exploring AI-driven development

Summary and Outlook

Zero Factory is an ambitious project exploring the form of software development through AI agent team collaboration. Through the collaboration of six specialized agents, clear workflow design, and cost optimization strategies, it demonstrates the future possibilities of AI-assisted development.

It is not just a tool but an idea: AI does not replace human developers but becomes a teammate. Each plays to their strengths to jointly produce better software.