Agentic AI and MCP for Product Managers
Lesson 2
Where the Seconds Go
Agents are slow in a way that single-step features are not, and the slowness compounds. Each pass through the loop adds a model call plus a tool call, and users feel every one of them.
Per Loop Pass | Typical | Over 8 Passes |
|---|---|---|
Model thinks and chooses a tool | 1.5-4s | 12-32s |
Tool executes | 0.1-2s | 1-16s |
Model reads the result | included above | - |
Total | 2-6s per pass | 20-50 seconds |
Forty seconds is a long time to stare at a screen with nothing happening. This is why Chapter 7 spends most of its length on what the user sees while waiting, it is not decoration, it's the difference between a feature people use and one they abandon on the second attempt.