DevSwarm: The Self-Evolving AI Software Agency

DevSwarm is an autonomous, event-driven multi-agent system that doesn't just write code—it architects its own workforce. Unlike static AI agents, DevSwarm performs a "Discovery Phase" to research modern best practices and dynamically hires specialized agents and creates custom tasks at runtime to build complex, multi-module software systems.

🧠 The "God Mode" Logic

DevSwarm operates on a recursive intelligence loop:

1. Reconnaissance: The Architect researches the user requirement via Google/Serper.
2. Self-Assembly: The system identifies gaps in its own "team" and dynamically instantiates new Agents (e.g., "Stripe Expert") and Tasks.
3. Sandboxed Execution: Code is written and then verified in a safe execution environment.
4. Self-Healing: If the QA agent detects a crash, the error log is fed back for an automatic refactor.

🚀 Key Features

• Dynamic Workforce Injection: Programmatically "hires" specialized agents based on project complexity.
• Hierarchical Supervision: Uses a Manager LLM (GPT-4o) to oversee Worker LLMs (GPT-4o-mini).
• Structured Output (Pydantic): All architectural blueprints are enforced via strict data models.
• Automated Testing: Includes a CodeExecutionTool to verify generated scripts before finalization.
• Cost Guardrails: Hard-coded logic limits the number of dynamic agents to prevent token-drain loops.

---

🛠️ Installation

1. Prerequisites

• Python 3.11+
• OpenAI API Key
• Serper API Key (for web research)

2. Setup

# Clone the repository
git clone https://github.com/Sama-ndari/dev-swarm-autonomous-agency.git
cd dev-swarm-autonomous-agency

# Install dependencies using uv
uv sync

3. Environment Configuration (.env)

Create a .env file in the root directory:

# Primary model for coding and documentation
MODEL=gpt-4o-mini
# High-reasoning model for management and architecture
MANAGER_MODEL=gpt-4o

OPENAI_API_KEY=sk-proj-xxxx
SERPER_API_KEY=xxxx

---

📂 Project Structure

dev_swarm/
├── src/
│   └── dev_swarm/
│       ├── tools/            # Custom logic: File Writers & QA Runners
│       ├── config/           # Agent & Task YAML templates
│       ├── crew.py           # The "Engine" (Agent Factory)
│       └── main.py           # The "Orchestrator" (Dynamic Loop)
├── project_output/           # THE FINAL PRODUCT: 
│   └── src/                  # Generated Source Code
│   └── README.md             # Generated Documentation

🎮 Usage

Run the swarm with a single command:

uv run crewai run

The agents will begin the discovery phase, create a blueprint.json, hire specialized agents, and build your project inside the project_output folder.

---

🛡️ Security & Guardrails

• Safe Execution: Code is executed in a subprocess/sandbox environment to protect the host machine.
• Resource Limits: The orchestrator enforces MAX_AGENTS and MAX_TASKS limits.
• Security Audit: A dedicated Sentinel agent scans for OWASP vulnerabilities in the generated output.

---

Created by Samandari - Advancing the frontier of Agentic Workflows.

---