Adaptive Model Router for Agents

Instead of one model per agent, a cheap Gemini Flash Lite classifier scores every step of a four-agent pipeline and routes it to GPT-5 mini, GPT-5 or GPT-5 Pro. By Partha Sarkar Project page: https://towardsdatascience.com/shipai/projects/adaptive-router ShipAI is a Towards Data Science project. You can learn more about it here: https://towardsdatascience.com/shipai/about Key takeaways. 1. A Gemini Flash Lite classifier scores every planned step 0–6 on complexity, reasoning and context, then picks a tier. 2. Context grows down the pipeline: the analyst reads all the research, the critic reads both, so later agents score higher. 3. On a cloud computing query, routing cuts estimated cost 40% against all-GPT-5 and 95% against all-GPT-5 Pro. Chapters: 0:00 Why fixed models fall short 1:32 Routing the researcher's steps 2:36 The analyst goes powerful 4:07 Cloud computing pipeline run 5:55 Autonomous weapons query 7:21 AI labor market query 8:11 How the router works