Self-Evolving Agent Ecosystem

Agent-Forge is an AI system that automatically creates, remembers, and improves specialized AI agents for any task. Describe what you need — the system builds an expert agent, saves it, and next time you need something similar, it's instant.

The system gets smarter with every task.

Every time you need AI to do something specific — write a cold email, analyze data, review code — you start from scratch. You write a prompt, iterate, maybe get a decent result. Tomorrow, you do it all over again.

Prompt engineering is a bottleneck. It's repetitive, inconsistent, and doesn't scale.

Agent-Forge replaces manual prompt engineering with an automated pipeline:

1. You describe the task — in plain language
2. System searches its skill library — has it solved something similar before?
3. If yes — loads the existing skill and executes instantly
4. If no — creates a new specialized agent with the right expertise, persona, and process
5. Saves everything — next time, it's a cache hit, not a rebuild
6. Learns — tracks what works, what doesn't, improves over time

• Auto-creation of specialized agents — each with a tailored persona, process, and quality standards
• Skill memory — agents persist across sessions, building a growing library
• Pattern library — 8 expert patterns (Analyst, Creator, Advisor, Processor, Orchestrator, Guardian, Teacher, Negotiator) matched to task type
• Self-improvement — system tracks quality scores and optimizes underperforming skills
• Model-agnostic — works with Claude, GPT, Gemini, or local models (Ollama)
• Tool integration (v2) — connects to 1,200+ tools via n8n for real automations, not just text

┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│   Your Task  │────▶│ Orchestrator │────▶│ Skill Search │
└──────────────┘     └──────┬───────┘     └──────┬───────┘
                            │                     │
                    ┌───────▼───────┐      Found? │
                    │    Execute    │◀─── YES ─────┘
                    │  with skill   │
                    └───────┬───────┘      NO ────┐
                            │                     │
                    ┌───────▼───────┐     ┌───────▼───────┐
                    │   Deliver     │     │    Prompt      │
                    │   Result      │     │   Architect    │
                    └───────────────┘     │  (creates new) │
                                         └───────┬───────┘
                                                 │
                                         ┌───────▼───────┐
                                         │  Save to Skill │
                                         │    Library     │
                                         └───────────────┘

• Node.js 20+
• pnpm
• Docker (for full stack)

git clone https://github.com/turboMe/AGENT-FORGE.git
cd AGENT-FORGE
pnpm install
cp .env.example .env
# Add your API keys to .env

# Start all services
docker compose up -d

# Run the API
pnpm dev

curl -X POST http://localhost:3000/task \
  -H "Content-Type: application/json" \
  -d '{"task": ""}'

Note: Agent-Forge is in early development.

packages/
├── api/              # REST API (Fastify)
├── orchestrator/     # Task routing and coordination
├── skill-library/    # Skill storage, search, indexing
├── llm-gateway/      # Multi-model LLM abstraction
├── memory/           # Persistent memory (Letta)
└── prompt-architect/ # Automatic agent/skill creation

Tech stack: TypeScript, Node.js, Fastify, MongoDB Atlas, Letta (MemGPT), Docker, Google Cloud Run.

• Core orchestration engine
• Prompt Architect — automatic skill creation
• Skill library with search and indexing
• Multi-model LLM gateway (Claude, GPT, local models)
• REST API with authentication
• n8n integration — tool-equipped automations
• Web UI — chat interface, skill browser, workflow dashboard
• Multi-agent collaboration
• Full offline mode with local models (Ollama)

_{Built in Reykjavik, Iceland 🇮🇸}