AgentForge is an experimental multi-agent AI programming system that simulates a structured software engineering team using Large Language Models (LLMs).
Instead of generating code with a single prompt, AgentForge introduces:
- Role-based AI agents
- Iterative collaboration cycles
- Controlled orchestration flow
- Architecture-level oversight
- Structured memory sharing
The project explores how AI agents can collaborate in a deterministic, system-governed workflow rather than operating as isolated prompt responses.
AgentForge simulates a real-world engineering team composed of three specialized agents:
| Agent | Role | Responsibility |
|---|---|---|
| Junior Developer | Implementation | Writes initial solution based on requirements |
| Senior Developer | Code Review | Reviews output, suggests improvements, identifies issues |
| Platform Engineer | Architecture & Oversight | Enforces best practices, validates structure, finalizes delivery |
Each session follows an orchestration cycle:
- User submits a task (e.g., "Build a scalable todo app")
- Junior agent generates initial implementation
- Senior agent reviews and refines
- Platform Engineer enforces architectural standards
- Iterations continue until convergence criteria are met
- Final output is delivered to the user
The focus is not just code generation — but controlled multi-agent collaboration.
- Explore deterministic orchestration in LLM systems
- Model structured collaboration instead of free-form chat
- Introduce role isolation and responsibility boundaries
- Experiment with iteration control and convergence logic
- Simulate AI-native development workflows
This is not a toy chatbot.
It is an exploration of AI-driven system design patterns.
- Controlled iteration loop
- Role-based prompt isolation
- Shared session memory
- Deterministic execution boundaries
- Clear agent responsibility separation
Future improvements include:
- Token usage tracking
- Observability & logging layer
- Failure recovery mechanisms
- Evaluation heuristics for convergence
- Cost monitoring
- React 18 + TypeScript
- Vite
- TanStack React Query
- Tailwind CSS
- shadcn/ui
- React Hook Form + Zod
- Python 3.11+
- FastAPI
- OpenAI API
- Pydantic
frontend/
src/
components/
features/
session-chat/
components/
hooks/
utils/
session-create/
hooks/
lib/
pages/
shared/
backend/
docs/
- Component → Hook → Utility separation
- Feature-based modular structure
- Barrel exports for clean imports
- Single Responsibility Principle
- Type-safe theming
- Controlled state management
The interface uses a retro pixel-art aesthetic with neon color tokens:
| Color | Token | Usage |
|---|---|---|
Pink #ff3388 |
primary |
Platform Engineer |
Cyan #00FFDD |
secondary |
Junior Developer |
Green #33FF66 |
accent |
Senior Developer |
Utility classes include:
pixel-borderneon-glow-primarytext-glow-secondary
- Node.js 18+
- npm or yarn
cd frontend
npm install
npm run devApplication runs at:
http://localhost:8080
| Method | Endpoint | Description |
|---|---|---|
| POST | /sessions | Create session |
| GET | /sessions/:id | Retrieve session |
| POST | /sessions/:id/iterate | Execute next orchestration cycle |
- Deterministic orchestration engine refinement
- Token usage tracking per agent
- Logging & observability layer
- Retry & failure recovery logic
- Agent evaluation heuristics
- Full FastAPI backend implementation
- Unit & integration testing
Contributions are welcome.
Please:
- Follow existing architecture patterns
- Maintain separation of concerns
- Add documentation where needed
- Run linting before commits
- Use clear, conventional commit messages
Beginner-friendly issues will be tagged as good first issue.
Built and maintained by Hasnaat Iftikhar
LLM Systems Engineer focused on:
- Multi-agent orchestration
- AI-native development workflows
- Structured LLM system design
🌐 Website: https://hasnaat.work
📧 Email: hasnaatfreelancing@gmail.com
Open to:
- AI infrastructure collaboration
- Multi-agent systems research
- Remote platform engineering opportunities
MIT