Infosys Agentic Foundry

Infosys Agentic Foundry: Open-Source Framework for Customizable Pro-Code AI Agents

Agentic Foundry is a robust, open-source platform enabling developers to build, configure, and deploy customizable AI agents with minimal coding. It streamlines agentic workflow development through modular templates, tool integration, and scalable deployment. Agents can be exported as standalone code or grouped for embedding into other workflows or hosted applications. The Infosys Agentic Service is deployed on Azure Kubernetes Service (AKS), ensuring enterprise-grade scalability and reliability.

Key Features & Capabilities

1. Simplified Agent & Tool Management

• Pro-Code Agent Creation: Automates setup using tool logic and workflow definitions, reducing development time.
• Agent & Tool Lifecycle Management: Seamless onboarding, updating, and removal within the framework.
• In-Platform Customization: Modify tools, workflows, and control logic without external IDEs or redeployment.
• Reusable Components: Modular, self-contained agents/tools promote consistency and speed across projects.
• Tool Configuration: Agents use external functions (tools) for tasks like web search, calculations, or DB queries. Secure connection management avoids embedding credentials. Tools can be created using natural language queries.

2. Advanced Agent Templates

• ReAct Agent: Uses reasoning traces and iterative tool execution for precise single-task operations.
• Multi-Agent System: Based on Planner-Executor-Critic model. Planner creates step-by-step plans, Executor performs tasks, Critic scores outputs. Supports automated and Human-in-the-Loop (HITL) modes.
• Meta-Agent: Acts as Supervisor, orchestrating ReAct, Multi-Agent, or hybrid workers. Dynamically selects agents, aggregates responses, and delivers final output.

3. Human-in-the-Loop (HITL) & Feedback Mechanisms

• HITL Integration: Manual checkpoints allow human oversight during critical decision-making steps.
• Feedback-Driven Learning: Structured/unstructured feedback loops improve agent performance over time. Admins manage feedback via a dedicated interface.
• Tool Interrupt: Users can review, modify, and approve tool calls step-by-step, enhancing transparency and control.

4. Robust Evaluation Frameworks

• Ground Truth-Based Evaluation: Compares agent outputs with golden datasets using metrics like:
  • SBERT Similarity
  • TF-IDF Cosine Similarity
  • Jaccard Similarity
  • ROUGE (ROUGE-1, ROUGE-L)
  • BLEU Score
  • Sequence Match Ratio
  • Exact/Fuzzy Match
  • Optional LLM Score
  • Includes LLM-based summary and downloadable report.
• LLM-as-a-Judge Evaluation: Qualitative scoring using LLMs without predefined answers. Metrics include:
• Tool Utilization Efficiency: Tool selection accuracy, usage efficiency, precision, success rate.
• Agent Efficiency Score: Task decomposition, reasoning relevance, coherence, robustness, groundedness, fluency.
• Results visualized via Admin Dashboard with filters (agent name, type, model, score threshold).

5. Enterprise-Grade Observability & Deployment

• Telemetry & Monitoring: Uses Open Telemetry, Arize Phoenix, and Grafana for tracing, observability, and real-time dashboards. Tracks LLM inputs/outputs, prompts, errors, and agent states.
• CI/CD Pipeline: Automates Docker image creation and deployment across containerized environments and Kubernetes clusters. Ensures consistency and reduces manual errors.
• Azure Kubernetes Deployment: Containerized via Docker and deployed on AKS. Supports auto-scaled microservices for agility and fast recovery.

6. Versatile Agent Offerings & Management

• Agent-as-a-Service: Framework-based solution for configuring agents/tools with conversational interfaces.
• Vertical Agents: Domain-specific agents for Finance, Healthcare, Insurance, Retail, Communication, and Manufacturing.
• Horizontal Agents: Cross-domain functionalities like Email, File Search, Agentic RAG, and SDLC.
• User & Admin Management: Admin Screen with Role-Based Access Control (RBAC). Roles include Admin, Developer, and User, each with specific permissions for registration, onboarding, feedback approval, and evaluation.