DevConf.IN 2026

Let’s be honest: Agentic AI works beautifully in 30-second demos, but often falls apart the moment you push it into production. We’ve all seen it: agents that get confused by their own context windows, use a sledgehammer (Reasoning Models) to crack a nut (simple lookups), or hallucinate tool parameters because they were overwhelmed by JSON schemas. The problem isn't usually the model's IQ but the Agentic system architecture.

In this session, we will move beyond basic Agentic patterns to explore the system architecture required for robust enterprise agents. Drawing on recent, cutting-edge research, we will dissect why a single "God Model" fails and how a Mixture-of-Agents (MoA) approach succeeds.

We will explore:

• The "When to Reason" Problem: How to use Semantic Routing to classify user intent, dynamically routing queries to the most cost-effective model (e.g., routing logic puzzles to reasoning models vs. factual lookups to standard LLMs) to slash latency and token costs.  

• The "Junk Drawer" Problem: Why dumping 90+ tools into a context window breaks agent performance. This includes using a Mixture-of-Tools strategies, where we use diverse, expert agents with specialized tool access rather than one overloaded generalist.  

• The "Inner Monologue": How to implement Metacognition and governance loops in integration with the Model Context Protocol (MCP). We will demonstrate how agents can "think about their thinking," self-correcting their plans and validating facts before presenting an answer to the user.  

Stop building black boxes. Come learn how to architect agents that are auditable, efficient, and actually know which tool to use.