· Nicholas Nadeau · Speaking · 2 min read
Agentic AI Is a Routing Problem, Not an Intelligence Problem
A talk on why agents fail in production and how embedding-based routing, expertise graphs, and the abstention pattern build resilient agentic systems.
I gave a talk at GenAI Montreal’s special ConFoo edition meetup at Hotel Bonaventure in Montreal. The event brought together Montreal’s software development, Cloud-Native, AI, and CTO communities for a multi-community breakout evening.
The Core Argument
Everyone is racing to make agents smarter: bigger models, more tools, more reasoning. But agents that survive production aren’t the smartest ones. They’re the ones with the best routing.
Agents don’t fail because they can’t reason. They fail because they don’t know what they don’t know.
Key Patterns
Expertise Graphs Over Chains of Thought
Most agent architectures have one model doing everything: routing, reasoning, and responding. It’s like making the surgeon also do triage, billing, and physical therapy. Not because they can’t, but because they shouldn’t.
Instead of chains, I proposed using expertise graphs where queries are routed via embedding similarity, a nearest-neighbor lookup, not a reasoning task. This is semantic dispatch: deterministic, parallelizable, and scaling to thousands of specialists where LLM-based tool selection breaks down.
Agents That Know What They Don’t Know
The most underrated capability in agentic AI is not knowing the answer and saying so. The abstention pattern gives every specialist a confidence threshold. Below it, the agent hands off rather than guesses. A well-calibrated small model beats an overconfident large model in production.
Budget-Aware Routing
Every agent call costs tokens, time, and money. The routing layer should be budget-aware: route simple tasks to cheap, fast specialists and allocate more resources for complex ones. This is operations engineering, not AI research, and it’s the difference between a demo and a product.
Takeaways
- Stop asking your LLM to route. Use embeddings. Routing is a similarity problem, not a reasoning problem.
- Build agents that can say “I don’t know.” Calibration is a feature, not a bug.
- Think in graphs, not chains. Pipelines break. Graphs adapt.
Stop making your AI smarter. Start making it honest about what it doesn’t know.
