To summarize:

1. version, review and test AI features

2. sandbox it

3. use quotas/rate limits

4. integrate AI in platform-agnostic ways (don't tie your success to one lab)

5. you'll save money paying for the best models

6. don't incentivize lines of code or token spend

7. train engineers to work with agents

8. reward employees for "impact"

9. don't cut jobs unless you've proven you can do more with less

I will personally do anything to avoid working with agents, so I have that conflict with point 7, but that aside, this is refreshingly sane coming from an AI company. AS to why I'd avoid agents; I personally find the process of reviewing AI-generated code to be far less educational and informative than starting with LLM suggestions and writing the code myself.

I think I could maybe use the article's suggestions as a way of determining if a business is well-run. The author's suggestions are only really leverageable by a business with strong direction.

I think it has now been said in many other threads or forums. But it is a good summary.