Dark Agent
A semi-autonomous AI process that takes real actions on real systems but is not treated like a production system: no owner, no logs, no policy, no tests.
A dark agent is any AI-powered process that can take actions, touches real data or systems, and is not treated like a first-class production system.
Why it matters
Individually each one feels harmless. It is just a helper bot, just a codegen script, just a workflow. Collectively they form a shadow workforce of non-human employees you never onboarded, never gave a job description, and never pen-tested. When one breaks, nobody gets paged, because nobody owns it.
In practice
You probably have dark agents right now. An LLM automation touching production data. An AI feature that can send email, modify code, or call external APIs. A workflow where you could not replay exactly what the AI did last Tuesday. If you cannot name the owner who gets paged when it fails, it is dark.
Where it shows up in my work
This is the opening move of the Agent Civilization Field Manual. Naming dark agents is how you stop pretending the AI is a feature and start treating it like the unvetted employee with root access that it actually is.