Your Foundation to Create Enterprise Agentic Platforms and Agents
📖 Documentation · 🚀 Quick Start · 🏗️ Architecture · 📋 Features · 📝 Release Notes
Infosys Agentic Foundry (IAF) is a comprehensive, open-source framework designed to empower developers to create, configure, and deploy customizable AI agents with minimal coding effort. It serves as a complete AI Agent Operating System — providing everything from visual agent design to production-grade deployment across multiple cloud providers.
The platform supports the full agent lifecycle:
- Design Time — Create, configure, and perfect AI agents with visual builders, 8 templates, workflow orchestration, and evaluation frameworks
- Runtime — Execute agents at scale with multi-LLM gateway, Kafka message queues, hyper-scale storage, and full observability
IAF is successfully deployed on Azure Kubernetes Service (AKS), AWS, and GCP, ensuring enterprise-grade scalability and reliability.
| Metric | Value |
|---|---|
| 🤖 Agent Templates | 8 (React, React-Critic, Planner-Executor, Planner-Executor-Critic, Planner-Meta, Meta, Hybrid, Skill) |
| 🏗️ Architecture Layers | 8 (Experience → Orchestration → Context → Reasoning → Tools + Cross-cutting) |
| ☁️ Cloud Providers | 3 (AWS, Azure, GCP) |
| ⚡ Inference Engines | 3 (LangGraph, Google ADK, Python-based) |
| 🔐 User Roles | 5 (Super Admin, Admin, Developer, User, Auditor) |
| 🔌 API Endpoints | 22+ |
| 🎯 Total Features | 97 (46 Functional + 51 Non-Functional) |
┌───────────────────────────────────────────────────────────────────────────────────────┐
│ INFOSYS AGENTIC FOUNDRY (IAF) PLATFORM ARCHITECTURE │
├───────────────────────────────────────────────────────────────────────────────────────┤
│ ┌─────────────────────────────────────────────────────────────────────────────────┐ │
│ │ Layer 0: EXPERIENCE — Agent Studio │ Chat UI │ MCP Console │ Ambient Inbox │ │
│ └─────────────────────────────────────────────────────────────────────────────────┘ │
│ ▼ │
│ ╔═════════════════════════════════════════════════════════════════════════════════╗ │
│ ║ Layer 1: AGENT ORCHESTRATION ★ THE HEART ★ ║ │
│ ║ React │ React-Critic │ Planner-Exec │ Plan-Exec-Critic │ Meta │ Hybrid │ Skill ║ │
│ ╚═════════════════════════════════════════════════════════════════════════════════╝ │
│ ▼ ▼ │
│ ┌────────────────────────────┐ ┌────────────────────────────────────────────┐ │
│ │ Layer 2: CONTEXT & MEMORY │ │ Layer 3: DECISION & REASONING (LiteLLM) │ │
│ │ SafeKernel │ /shared/ │ │ Azure │ OpenAI │ Ollama │ Google ADK │ │
│ │ /memory/ │ /workspace/ │ │ SBERT │ BGE │ CoT │ ReAct │ ToT │ │
│ └────────────────────────────┘ └────────────────────────────────────────────┘ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────────────────────────┐ │
│ │ Layer 4: TOOL EXECUTION — Code Executor │ MCP Registry │ Python Tools │ APIs │ │
│ │ Backends: Subprocess │ Docker │ nsjail │ Async + Cache │ │
│ └─────────────────────────────────────────────────────────────────────────────────┘ │
├───────────────────────────────────────────────────────────────────────────────────────┤
│ Cross-cutting: Evaluation │ Feedback Learning │ Export │ Kafka MQ │ Observability │
│ Security: RBAC (5 Roles) │ Dept Isolation │ VAULT │ SANDBOX │ AUDIT LOGGING │
└───────────────────────────────────────────────────────────────────────────────────────┘
Core Architectural Concept — The Controller Pattern:
Request → Orchestrator (L1) → Queries Context (L2) → Invokes Reasoning (L3) → Executes Tools (L4) → Response
| Template | Pattern | Description |
|---|---|---|
| React | Reason → Act | Single agent with step-by-step reasoning and tool execution |
| React-Critic | Reason → Act → Validate | React + built-in self-critique for higher accuracy |
| Planner-Executor | Plan → Execute | Separates planning from execution with replanning support |
| Planner-Executor-Critic | Plan → Execute → Validate | Three-stage cycle with quality validation |
| Planner-Meta | Plan → Orchestrate | Advanced orchestrator with multi-prompt planning and delegation |
| Meta | Orchestrate → Delegate | Supervisor agent coordinating multiple worker agents |
| Hybrid | Pure Python | Framework-free agent with native planning and execution |
| Skill | Declarative | SKILL.md-based declarative agents with zero-code creation |
- Visual DAG Builder — Drag-and-drop workflow designer for agent chaining
- Sequential & Parallel Execution — Run agents in series or concurrently
- Conditional Branching — Route data based on agent outputs
- Human-in-the-Loop (HITL) — Plan approval with feedback at critical decision points
- Agent Pipelines — Chain multiple agents into deterministic, reusable workflows
- Python-to-MCP server conversion automation
- MCP server CRUD with registry management
- External MCP URL support with custom headers
- Real-time tool discovery with enterprise security and audit logging
- Tool & MCP export/import across environments
- LiteLLM Proxy — Unified interface for Azure OpenAI, OpenAI, Ollama, and Google ADK (Gemini)
- Token Tracking — Per-request token usage analytics
- Cost Calculation — Model-based cost tracking per user, agent, and department
- Load Balancing & Fallback — Automatic failover between LLM providers
- Custom Model Support — Bring your own models with configurable endpoints
| Method | Description |
|---|---|
| Ground Truth | Automated comparison against golden datasets (SBERT, ROUGE, BLEU, Jaccard, TF-IDF, Exact/Fuzzy Match) |
| LLM-as-a-Judge | Multi-dimensional scoring without predefined answers |
| Consistency Testing | Temporal consistency across repeated queries |
| Robustness Testing | Adversarial input evaluation |
| Phoenix Integration | Trace visualization and debugging |
- Semantic Memory — Persistent cross-session fact storage via Redis and PostgreSQL
- Episodic Memory — Few-shot learning from past conversations using similarity scoring
- Custom Knowledge Bases — Upload documents (PDF, TXT) for domain-specific intelligence
- Feedback-Driven Learning — Continuous improvement loop: User Feedback → Lesson Extraction → Admin Approval → Knowledge Update → Agent Improvement
- 5-Role RBAC — Super Admin, Admin, Developer, User, Auditor with department-based isolation
- Secrets Vault — Master key management for API keys, URLs, and credentials
- Rate Limiting — Per-user sliding window protection
- Audit Logging — Complete operation tracking
- Sandboxed Execution — Isolated tool execution via Docker/nsjail
- JWT Authentication — Secure Bearer token authentication for all endpoints
- OpenTelemetry — Full distributed tracing integration
- Arize Phoenix — LLM observability with trace visualization
- Token Usage Reports — Per-user, per-agent, per-model analytics
- Response Metrics — Per-agent response time and performance tracking
- Grafana Dashboards — Real-time monitoring and alerts
- Standalone Agent Export — Export agents as independent Python packages with all dependencies
- GitHub Push — Direct repository push with tool versioning
- Blob Storage — Cloud storage for export artifacts (AWS S3, Azure Blob, GCP)
- Docker & Kubernetes — Production-ready containerization with AKS/EKS/GKE deployment
- Multi-Cloud Support — Deploy on AWS, Azure, or GCP with unified abstractions
| Feature | Description |
|---|---|
| Viber Agent | Conversational AI assistant that creates agents from plain descriptions — zero technical knowledge required |
| SSE Streaming | Real-time streaming of agent execution steps to the UI |
| Canvas Screen | Rich visualization of tables, charts, graphs, and images in chat |
| Prompt Optimization | Automated prompt evolution using Pareto sampling and LLM-as-judge scoring |
| Validators | Custom response validation logic with scoring and real-time feedback |
| Tool Interrupt | Review, modify, and approve tool calls step-by-step before execution |
| Data Connectors | Connect to PostgreSQL, SQLite, MySQL, and MongoDB |
| Kafka Message Queue | Async tool/agent execution, batch processing, M2M communication |
| GZIP Compression | Optimized response payloads for performance |
| Google ADK Support | Full Google Agent Development Kit as an inference backend alongside LangGraph |
- Python 3.10+
- Node.js 18+
- PostgreSQL & Redis
- Docker (optional, for containerized deployment)
IAF supports multiple deployment options:
| Platform | Guide |
|---|---|
| 🪟 Windows | Windows Setup |
| 🐧 Linux | Linux Setup |
| 🐳 Docker Compose | VM Docker-Compose |
| ☁️ Azure (AKS) | Azure Deployment |
| ☁️ AWS (EKS) | AWS Deployment |
| ☁️ GCP (GKE) | GCP Deployment |
Infosys-Agentic-Foundry/
├── Infosys-Agentic-Foundry-Backend/ # FastAPI backend (Port 8080)
│ ├── src/ # Core source code
│ ├── agent_worker/ # Kafka agent worker
│ ├── tool_worker/ # Kafka tool worker
│ ├── knowledgebase_server/ # Knowledge base service
│ └── Export_Agent/ # Agent export module
├── Infosys-Agentic-Foundry-Frontend/ # React frontend (Port 3000)
│ └── src/ # React components & pages
├── IAF-Litellm-Server/ # LiteLLM proxy server
├── docs/ # MkDocs documentation
├── manifest_file/ # Kubernetes manifests
└── site/ # Built documentation site
Full documentation is available at https://Infosys.github.io/Infosys-Agentic-Foundry/
- Getting Started
- Architecture Overview
- Agent Design Patterns
- Agent Configuration
- Tool Configuration
- MCP Registry
- Evaluation Framework
- Admin Screen
- RBAC & Security
- Telemetry & Monitoring
- Installation Guides
💬 User Feedback → 🧠 Lesson Extraction → ✅ Admin Approval → 📚 Knowledge Update → 🚀 Agent Improvement
The platform continuously improves agent performance through structured feedback collection, automated lesson extraction, admin-controlled approval workflows, and knowledge base updates that feed back into agent behavior.
We welcome contributions! Please see our contribution guidelines and ensure your code follows the project's standards.
This project is licensed under the Apache License 2.0.
Infosys Agentic Foundry (IAF) — Part of Infosys Topaz
Enterprise-Grade AI Agent Operating System • V1.9.0 • May 2026