Microsoft Developer: Cut AI Agent Costs with TokenOps

AI agent costs don't grow in step with the work. As an agent takes more turns, context piles up and uncached input costs rise on a quadratic curve: in Gareth Bland's example, an agent that takes four times as many turns costs sixteen times as much. Engineering your context to use cached input can bring that curve down to a tenth of the cost.



Gareth Bland, Chief Data Scientist at Microsoft, explains TokenOps: the practice of deciding where token spend will move the needle most. He covers why reasoning makes output tokens the expensive part, how to measure agentic efficiency with three ratios (autonomous completion rate, token efficiency ratio, and value generation ratio), and how an S-curve shows whether your team is underinvesting, getting real leverage, or gold-plating features nobody asked for.



Resources

- Unit Economics of Agentic AI https://aka.ms/tokenomicon-2026

- Maximize ROI from AI https://azure.microsoft.com/solutions/maximize-roi-from-ai

- More Azure resources! https://azure.com/AzureEssentials



Connect

- Gareth Bland https://www.linkedin.com/in/gareth-bland/



Chapters

0:00 What TokenOps means for AI agent spending

0:35 Why output tokens cost more than input tokens

1:02 Why AI agent costs grow faster than the work

1:50 How caching brings AI agent costs down

2:05 How to measure agentic efficiency

3:05 How to know where your next token dollar belongs