Infosys Agentic AI Foundry, part of Infosys Topaz, is a suite that helps enterprises build reliable agents using various design patterns. Agents can be deployed as-is or with custom UX. Built for enterprise-grade reliability, it enables businesses to confidently reimagine their processes.
For more details check the README.md inside the Frontend or Backend folder.
- Tool Management: Introduced functionality to create, onboard, and manage custom tools (Python functions) for LLM agents, making them reusable. Users can write custom logic for tools to access uploaded file content.
- Agent Templates: Released initial agent templates including:
- React Agent for single-task accomplishment.
- Multi-Agent system using a Planner-Executor-Critic framework, with an optional "Plan Verifier" mode.
- Memory & Interaction: Implemented simple conversation memory for agents and enabled real-time interaction via chat for inference and task execution.
- File Integration: Added support for file uploads, allowing agents to utilize file content through integrated tools.
- UI Enhancements: Introduced live tracking for agent activities and provided access to documentation via a Help button.
- Admin Panel: Launched a comprehensive Admin Panel with features for:
- User registration, User profile updates, and a recycle bin for deleted items.
- Evaluation Dashboard displaying evaluation results, tool utilization, and agent efficiency scores.
- Metric management: Allows admins to trigger evaluations using the "LLM as a Judge" approach.
- Feedback Management: Admins can review and approve user feedback before it is used for agent training.
- New Agent Templates:
- Meta Agent (Supervisor Agent): Supervisor agent template which delegates tasks to worker agents.
- Planner Meta Agent: Two-stage workflow—first, a planner node generates a step-by-step plan; then, each step is executed sequentially by the meta/supervisor agent.
- Planner-Executor Agent: Template similar to Multiagent without Critic, with and without "Plan Verifier".
- React Critic Agent: Template similar to React introduced in version 1.0.0, with the addition of Critic.
- Agent Creation & Export: Users can export the code for any agent, along with the tools used by agents, to run locally and modify as needed.
- Feedback Learning: Agents (React and Multi-Agent) learn from user feedback to improve future responses, subject to admin approval.
- System Enhancements:
- Added tool interrupt capability for both React Agent and Multi-Agent workflows, allowing dynamic intervention and correction during tool execution.
- Refactored core functions using
async/awaitfor scalability, and improved feedback handling in the multi-agent system.
- Telemetry & Tracing:
- Enabled OpenTelemetry to monitor agent activities and system performance for improved diagnostics, analytics, and framework-level debugging.
- Integrated Arize Phoenix for detailed agent tracing and monitoring.
- Evaluation Metrics: Introduced evaluation metrics to assess the performance and efficiency of AI agents, focusing on the "LLM as a Judge" approach.
- UI Updates: Added new UI pages for updating users and managing the recycle bin.
- Code Architecture Modernization: Refactored code using Object-Oriented Programming (OOP) principles, introduced modular service layers (AgentService, ToolService, and TagService), and implemented a repository pattern with database connection pooling for improved performance.
- Enhanced Agent Templates: Streamlined Multi-Agent templates, adding a Planner & Executor configuration and a React & Critic setup for focused evaluation.
- Advanced Features:
- Introduced Ground Truth Based Evaluation for precise agent output assessment.
- Integrated Knowledge Base support for React agents.
- Database & Connectivity: Enhanced Data Connector functionality to support multiple databases (SQLite, PostgreSQL) and updated existing functions for PostgreSQL migration.
- Security & Management (Vault): Developed a centralized Secrets Handler module for secure management of sensitive information, including public and private secrets.
- Monitoring Improvements: Enhanced telemetry with trace ID fixes and session tracking, and modularized chat-related logic into dedicated service layers.
- Admin & UI Refinements: Improved Recycle Bin logic, added API endpoints for listing markdown files, introduced new UI pages for Ground Truth and Secrets, and resolved various defects across admin screens and tool mapping.
- Continued Modularization: Further modularized key functionalities, including feedback learning and evaluation metrics, for improved maintainability and integration into agent export.
- Configuration: Introduced configurable CORS origins and SBERT model paths via environment variables.
- System Stability: Addressed public/private key management issues.
- Development & Integration: Performed legacy code cleanup and unified API access through FastAPI integration.
- UI & Agent Fixes: Released Chat UI Version 2.0 (Inference page design with new look), added a new Data Connectors page, corrected API endpoints due to modularity changes, and resolved issues on agent screens related to Meta/Planner Meta Agents, System Prompts, and Evaluation metrics menu item changes.
- Inference Output Format Restructuring: Restructured API endpoints' inference output to include a list of all tools invoked during the agent inference call along with their corresponding outputs.