Each post explains one idea for a technical investor or a platform engineer, and links to the published result behind it. Every result figure in the posts is read from a published file when the pages are built, and is marked with a dotted underline. Subscribe by feed.
What crash recovery means for AI agents that change real files and databases
An agent that dies halfway through a multi-step change can leave a mess, and a journal on disk is the usual way to make that change all or nothing.
Why a set of AI agent tools can be unsafe even when every pair is safe
Checking agent tools one at a time, or two at a time, can pass a combination of three that leaks a secret.
Why testing a system at crash points you chose proves less than it seems
A recovery design that passes every crash point its authors chose can still fail at moments they did not think of.
How to approve AI agent tools as a set, not one at a time
Approving an agent's tools as a set means asking what the whole kit can do, and recording each approval so that it cannot be reused by accident.