Version: 1.0
AgentForge is a production-oriented AI Agent Platform designed to serve as a foundation for building intelligent assistants, automation systems, infrastructure operators, and knowledge workers.
Unlike traditional Retrieval-Augmented Generation (RAG) systems that focus exclusively on document retrieval and question answering, AgentForge treats RAG as a tool within a larger agent ecosystem.
The platform provides:
- Agent execution runtime
- Tool orchestration
- Memory systems
- Knowledge management
- Workflow execution
- Human approval gates
- Multi-provider LLM integration
- Observability
- Asynchronous task execution
- Infrastructure automation capabilities
AgentForge is designed to evolve from a single-agent platform into a distributed multi-agent system.
Modern AI systems require more than retrieval.
Organizations increasingly need agents that can:
- Understand goals
- Create plans
- Execute actions
- Use tools
- Access knowledge
- Maintain memory
- Collaborate with humans
- Automate workflows
AgentForge aims to provide a framework where new capabilities can be added as tools without modifying the core runtime.
Agents never directly interact with external systems.
Every capability is implemented through a tool interface.
Examples:
- RAG Tool
- Search Tool
- SQL Tool
- File Tool
- Kubernetes Tool
- Firecracker Tool
- Ceph Tool
- Email Tool
The runtime is responsible for:
- Planning
- Tool selection
- Context management
- Memory access
- Execution orchestration
- Result aggregation
The runtime never directly queries databases.
All persistence is accessed through service interfaces.
Long-running operations are delegated to Celery workers.
Examples:
- Document ingestion
- Embedding generation
- Knowledge indexing
- Agent execution
- Workflow execution
Any LLM provider should be replaceable.
Examples:
- OpenAI
- Anthropic
- Ollama
- Local vLLM
- Azure OpenAI
User Request
↓
API Layer
↓
Agent Runtime
↓
Planner
↓
Tool Registry
↓
Tool Execution
↓
Memory Integration
↓
Response Generation
↓
User
Backend
- Python 3.12+
- Django
- Django REST Framework
Asynchronous Processing
- Celery
- Redis
Database
- PostgreSQL
Vector Storage
- pgvector
LLM Providers
- OpenAI
- Anthropic
- Ollama
- vLLM
Monitoring
- Prometheus
- Grafana
- OpenTelemetry
Deployment
- Docker
- Docker Compose
- Kubernetes
agentforge/
manage.py
config/
apps/
core/
providers/
workers/
tests/
docs/
Responsible for executing agents.
Responsibilities:
- Context management
- Tool orchestration
- Memory integration
- Plan execution
- Result aggregation
The runtime is the heart of the platform.
Converts goals into executable plans.
Example:
Goal:
"Find all invoices from April and summarize spending"
Plan:
- Search documents
- Retrieve invoices
- Extract text
- Aggregate totals
- Generate summary
Maintains available tools.
Responsibilities:
- Tool registration
- Discovery
- Validation
- Permissions
Stores active conversations.
Technology:
- Redis
Used for:
- Conversation state
- Agent state
- Active workflows
Stores persistent knowledge.
Technology:
- PostgreSQL
- pgvector
Used for:
- Agent memories
- Knowledge bases
- Historical interactions
The knowledge subsystem provides:
- Document ingestion
- Chunking
- Embedding generation
- Indexing
- Retrieval
Supported formats:
- Markdown
- DOCX
- TXT
- HTML
RAG is implemented as a tool.
Pipeline:
Upload
↓
Chunk
↓
Embed
↓
Store
↓
Retrieve
↓
Context Assembly
↓
LLM
AgentForge implements hybrid retrieval.
Methods:
Keyword-based ranking.
Advantages:
- Precise matching
- Fast retrieval
Semantic similarity search.
Advantages:
- Concept matching
- Context awareness
Combines both methods.
Result:
Higher retrieval quality.
Every tool must implement:
Tool Metadata
- Name
- Description
- Permissions
- Input Schema
- Output Schema
Execution Interface
execute()
Validation Interface
validate()
Provides document retrieval.
Provides internet search.
Provides filesystem access.
Provides database querying.
Provides workflow execution.
Provides scheduled tasks.
Future capabilities:
- VM creation
- VM deletion
- VM lifecycle management
- Container deployment
- Container monitoring
- Storage operations
- Volume management
- Cluster operations
- Deployment management
Workflows allow agents to execute repeatable processes.
Example:
User Onboarding
- Create account
- Send email
- Provision resources
- Notify team
Some operations require approval.
Examples:
- Infrastructure changes
- Database migrations
- VM deletion
Agent execution pauses until approval is granted.
Create Agent
List Agents
Update Agent
Delete Agent
Run Agent
Get Execution Status
Get Execution Logs
Get Execution Results
Upload Document
Delete Document
Reindex Document
Search Knowledge Base
List Tools
Enable Tool
Disable Tool
View Tool Metadata
ingestion
chunking
embedding
indexing
agent_execution
tool_execution
workflow_execution
maintenance
Authentication
- JWT
Authorization
- RBAC
Roles
- Admin
- Operator
- Developer
- Viewer
Metrics
- Request latency
- Tool execution time
- Retrieval performance
- Agent runtime performance
Logs
- Structured JSON logs
Tracing
- OpenTelemetry
Unit Tests
- Services
- Tools
- Providers
Integration Tests
- Retrieval
- Agent Runtime
- API
End-to-End Tests
- Full execution paths
Phase 1
Core Platform
- Agent runtime
- Tool registry
- Knowledge management
- Hybrid retrieval
Phase 2
Advanced Agents
- Planning
- Reflection
- Memory systems
Phase 3
Multi-Agent System
- Agent collaboration
- Agent messaging
Phase 4
Infrastructure Operations
- Firecracker integration
- Ceph integration
- Kubernetes integration
Phase 5
Enterprise Features
- Multi-tenancy
- Audit trails
- SSO
- Approval workflows
AgentForge should evolve into a platform where organizations can create specialized AI workers that interact with knowledge, software systems, infrastructure, and business processes through a unified agent runtime.
The ultimate objective is to provide a flexible foundation for building operational AI systems rather than isolated chatbots.