I designed nine AI employees. None of them has worked a day.
A mid-build status report instead of a launch announcement — nine scoped agent roles, one real report, and zero unattended runs so far.
Topic
Security & Data Exposure
Who can actually reach your data, not who is supposed to — reviewing AI-written code and drawing boundaries that are enforced rather than described.
3 articles
What my AI assistant is not allowed to do
The interesting engineering in an AI agent is not what it can do. It is what it cannot — and those limits have to be code, not prompt text.
"Enabled" is a claim, not evidence
Software you depend on will not fail loudly — it reports that it is fine. The green lights in my own systems that meant nothing, and what to test instead.