Geek Out Time: Supercharging Multi-Agent AI: Hybrid Routing with AG2(formerly Autogen)

In an earlier post, I built a career‑coaching chatbot with four domain agents — Passion, Obstacle, Balance, Values ( https://medium.com/the-constellar-digital-technology-blog/geek-out-time-building-an-interactive-career-coach-with-autogen-groupchat-37ddbcc38d2e ) . Useful, but a bit noisy. This week I upgraded the orchestration so that only the right agents speak, and a LeadCoachAgent turns their ideas into a crisp plan.

How routing is done in real‑world agent stacks

In production multi‑agent systems, teams rarely rely on a single heuristic. At scale the common pattern is hybrid routing :

This layered approach keeps latency low (≈20 ms for 80 % of queries) while letting the LLM spend tokens only where it adds value. The HybridRouter you’ll see below is a distilled version.

This is an excerpt — the full article continues on Medium.