A 4-tier setup for cheaper, faster AI agents
AI agent work can be split across four s instead of sending every request to one powerful model. Most requests first go to a cheap, fast that classifies the task and decides whether it needs a stronger model. The main idea is that most agent calls do not need a ; they need quick routing first, then stronger reasoning only when the task is hard enough.
Matching the model level to the task level improved both cost and the feel of the agent loop more than choosing one smarter model for everything. Speed was the main for interactive agents. When a supervisor takes more than 10 seconds for each decision, the whole agent feels slow even if the answers are good.
When the decides in 2 to 5 seconds, the loop feels much more usable. The example stack uses as the , GLM-5.2 as the primary advisor, GLM-5.2 at maximum effort for , and Opus 4.8 only for sanitized high-stakes work.
Key points
- Most agent requests can start with a cheap .
- Expensive models are only used when the task needs stronger reasoning.
- Routing work by difficulty can reduce cost and improve the feel of the agent loop.
- A supervisor that takes over 10 seconds per decision can make an agent feel sluggish.
- The example stack has four roles: , advisor, , and premier high-stakes handling.