python run_hephaestus_dev.py --path /path/to/project

Get started with Hephaestus Dev →

The Problem I Kept Running Into

I was trying to build a system where AI agents could handle complex software projects. You know the kind: "Build me an authentication system with OAuth, JWT, rate limiting, and comprehensive tests."

Traditional agentic frameworks can branch and loop, but they have a limitation: every branch needs predefined instructions. You must write the task descriptions upfront for every scenario you anticipate.

But what about discoveries you didn't anticipate? When a testing agent finds an optimization opportunity, a security issue, or a better architectural pattern?

Here's what I tried instead: Define logical phase types that are needed to solve problems - like "Plan → Implement → Test" - and let agents create tasks in ANY phase based on what they discover.

What Actually Happened: A Branching Tree That Builds Itself

Instead of a rigid sequence, I set up phase types:

• Phase 1 (Analysis): Understanding, planning, investigation
• Phase 2 (Implementation): Building, fixing, optimizing
• Phase 3 (Validation): Testing, verification, quality checks

The key insight: Agents can spawn tasks in any phase they want.

A validation agent testing your auth system might discover an elegant caching pattern. Instead of being stuck (or following predefined branching logic you wrote), the agent:

1. Creates a Phase 1 investigation task: "Analyze auth caching pattern - could apply to 12 other API routes for 40% speedup"
2. Keeps working on their validation task
3. Another agent picks up the investigation task and explores it

The workflow just branched itself. Not because you predicted "if optimization found, spawn investigation task" - but because the agent discovered something worth exploring and had the freedom to create work for it.

This creates a branching tree of tasks that grows based on actual discoveries, not anticipated scenarios.

Let me show you what this looks like in practice:

Example: Building from a PRD

I give Hephaestus a product requirements document: "Build a web application with authentication, REST API, and a React frontend."

Phase 1 agent reads the PRD and identifies 5 major components:

1. Authentication system
2. REST API layer
3. React frontend
4. Database schema
5. Background workers

It spawns 5 Phase 2 tasks — one for each component. Now I have 5 agents building in parallel, each focused on one piece.

One of the Phase 2 agents finishes the REST API and spawns a Phase 3 validation task: "Test the REST API endpoints."

The Phase 3 agent starts testing. Everything passes. But then it notices something:

"The auth endpoints use a caching pattern that reduces database queries by 60%. This could speed up all API routes significantly."

Here's where it gets interesting.

The Phase 3 agent doesn't just log this observation and move on. It doesn't get stuck because there's no "investigate optimizations" in the workflow plan.

Instead, it spawns a new Phase 1 investigation task: "Analyze auth caching pattern — could apply to other API routes for major performance gain."

A new Phase 1 agent spawns, investigates the caching pattern, confirms it's viable, and spawns a Phase 2 implementation task: "Apply caching pattern to all API routes."

Another agent implements it. Another agent validates it.

The workflow just branched itself. No one planned for this optimization. An agent discovered it during testing and created new work to explore it.

Meanwhile, a different Phase 3 agent is testing the authentication component. Tests fail. So it spawns a Phase 2 bug fix task: "Fix auth token expiry validation — current implementation allows expired tokens."

The fix agent implements the solution and spawns Phase 3 retest: "Validate auth fixes."

What Just Happened?

Look at what emerged:

graph TB
    P1[Phase 1: Analyze PRD<br/>Creates 5 tickets] --> P2A[Phase 2: Build Auth]
    P1 --> P2B[Phase 2: Build API]
    P1 --> P2C[Phase 2: Build Frontend]

    P2B --> P3B[Phase 3: Test API]
    P3B -->|discovers optimization| P1New[Phase 1: Investigate Caching<br/>NEW BRANCH]
    P3B -->|testing continues| P3Done[API Validated]

    P1New --> P2New[Phase 2: Implement Caching]
    P2New --> P3New[Phase 3: Validate Optimization]

    P2A --> P3A[Phase 3: Test Auth]
    P3A -->|tests fail| P2Fix[Phase 2: Fix Auth Bug]
    P2Fix --> P3Retest[Phase 3: Retest Auth]

    style P3B fill:#fff3e0
    style P1New fill:#e1f5fe
    style P2Fix fill:#ffebee

This workflow built itself:

• Started with 1 analysis task
• Branched into 5 parallel implementation tasks
• One testing phase discovered optimization → spawned 3-phase investigation branch
• Another testing phase found bugs → spawned fix → retest loop
• All coordinated through Kanban tickets with blocking relationships

Why This Changes Everything

Traditional workflows: Predict every scenario upfront → rigid plan → breaks when reality diverges

Hephaestus approach: Define work types → agents discover → workflow adapts in real-time

The workflow adapts in real-time based on what agents actually discover, not what we predicted upfront.

The Semi-Structured Sweet Spot

Here's why this is "semi-structured" and why that matters:

Fully structured workflows (traditional frameworks):

• ❌ Require predefined prompts for every scenario
• ❌ Can branch/loop, but need fixed instructions for each path
• ❌ Must anticipate all discoveries upfront

Fully unstructured agents (chaos):

• ❌ No coordination
• ❌ Duplicate work
• ❌ Contradictory changes
• ❌ No clear success criteria

Semi-structured (Hephaestus):

• ✅ Phase definitions provide work type structure and guidelines
• ✅ Agents write task descriptions dynamically based on discoveries
• ✅ Kanban tickets coordinate work with blocking relationships
• ✅ Guardian monitoring ensures agents stay aligned with phase goals
• ✅ Workflow adapts to what agents actually find, not what you predicted

You get structure where it matters:

• Phase types define what kind of work is happening
• Done definitions set clear completion criteria
• Guardian validates alignment with phase instructions
• Tickets track dependencies and prevent chaos

And flexibility where you need it:

• Agents create detailed task descriptions on the fly
• No need to predefine every possible branch
• Discoveries drive workflow expansion in real-time
• New work types emerge as agents explore

🚀 Quick Start

Prerequisites

• Python 3.10+
• tmux - Terminal multiplexer for agent isolation
• Git - Your project must be a git repository
• Docker - For running Qdrant vector store
• Node.js & npm - For the frontend UI
• Claude Code, OpenCode, Droid, or Codex - CLI AI tool that agents run inside
• API Keys: OpenAI, OpenRouter, Anthropic (also supports: Azure OpenAI, Google AI Studio - see LLM Configuration)

Validate Your Setup (macOS)

Before starting, validate that everything is installed and configured correctly:

python check_setup_macos.py

This script checks:

• ✅ All required CLI tools (tmux, git, docker, node, npm, Claude Code)
• ✅ API keys in .env file
• ✅ MCP servers configured
• ✅ Configuration files and working directory
• ✅ Running services (Docker, Qdrant)