Agentic AI | Multi Agent Collaboration

Delivered as part of the FDP on "Orchestrating Intelligence: Multi-Agent Collaboration in Multimodal Generative AI Ecosystems" at Dayananda Sagar Academy of Technology and Management (DSATM), this 2.5-hour session served as the structural backbone for the program.

Defining the Agentic Boundary: Dr. Mahalingam distinguished agentic AI (goal-directed, tool-using, stateful, iterative, and bounded) from standard predictive ML and one-shot generative AI, establishing an honest threshold for when simpler workflows should be favored over complex agentic loops.

Lessons from Production: Analyzing a real-world outbound voice agent, the talk demonstrated how latency was optimized—not through better models, but through system engineering like turn-detection, full streaming, and unblocking the critical path.

System Breakdown & Topologies: The session detailed four primary systemic forces that break single agents (context dilution, tool-surface overload, lack of failure isolation, and mismatched model/latency tiers) and evaluated core topologies such as Supervisor/Orchestrator, Sequential Pipelines, Routers, and Blackboard architectures.

Protocols & 2026 Frameworks: The presentation mapped the 6-layer enterprise architecture, contrasting Anthropic's Model Context Protocol (MCP) for tool access with Google's Agent-to-Agent (A2A) protocol for peer delegation, while comparing modern frameworks like LangGraph, Microsoft Agent Framework, and Google ADK.

Failure Modes & Governance: Citing empirical MAST-14 failure categories and compounding error rates, the talk addressed key governance requirements, auditability, decision provenance, and compliance under India's DPDP Act 2023.

Academic Integration: Concluding with a 3-week course syllabus, student project ideas, a group design exercise, and a 10-point evaluation checklist, the session equipped faculty with practical tools to teach agentic AI architectures effectively.