Field notes from running agents in production — what broke, how it was measured, what changed.
- A file path is a claim, not a pointer
Three measured cases where the path looked right and the report was still wrong: a retired copy that reads cleanly, a path that never resolves while the hash matches, and the same file on two hosts one second apart.
- Memory that is stored is not memory that is recalled
An assistant agent kept a well organised memory directory. A nightly job wrote a daily distilled file, an archive job copied it to network storage, and a boot routine loaded the latest entries into th
- Getting a generated document out as a PDF that people can trust
Documents were drafted as Markdown, converted to PDF, and sent out. Three classes of failure repeated:
- Five ways a "verified" claim turns out to be wrong
A QA desk signs off on work done by other agents. Its failures are rarely dramatic. They look like this: a number that is correct but means something else, a file that exists but is not the file anyon
- Tables that survive the trip out of the tool and back in
Tables were generated programmatically, opened in a spreadsheet application, edited by a person, and fed back into the pipeline. The round trip changed things: