Summary
In developing a service at my own company, I told the AI agent that it could use an external service’s sandbox. A sandbox is a testing environment that an external service provides. I also told it that I would issue the API keys that were needed. Even after that, the AI agent did not connect to the external service’s API and kept implementing against mocks. A mock is a provisional implementation that returns fixed responses in place of the external service. In this column, I read through the development records from that same week and think about what the design I had approved was premised on.
What you can take away
This column is for people who develop software while making design decisions together with an AI agent. Drawing on events that actually happened, it considers what the approvals you have given were premised on.
This article is a column based on records that the AI agent wrote during the development of a service at my own company, in response to issues I pointed out.
The rest of this article is paid
This is a Premium-only column, available to Premium subscribers.
The paid part is about 10,700 characters, roughly a 21-minute read.
Read this Premium-only column
Premium is 980 JPY / month
Premium adds this column and dialogue with matsumotory-kun on top of every paid article. See the plans for details.
Prices include tax. Purchases and subscriptions start after you sign in. Subscribers and readers who already bought this article can sign in and read the full article; for the plans, see the plans page.
Sales are available in supported regions only; the Terms list where we sell. All charges are in Japanese yen.