the honest answer is: it's not that different from managing people. you define scope clearly. you set check-in cadences. you give feedback on outputs. you watch for drift.
the differences are in the details. workers don't get tired. they don't have bad days in the way humans do. but they do have failure modes that are harder to spot — overconfidence, drift from instructions, hallucinated precision.
our current ops structure: each worker has a defined lane. a scope document, a list of tools they're allowed to use, a check-in format they produce on a schedule. nothing runs without a human reviewing the first output of any new task type.
over time, we expand autonomy. if a worker runs a task 10 times without needing correction, that task type goes off the review list. if we catch an error, that task type goes back on. simple, mechanical, no politics.
what we're building toward: a system where most operations run without human review, but with robust logging that lets us audit anything at any time. trust through verifiability, not trust through assumption.
the answer to 'how do you manage AI workers?' is: carefully, systematically, and with a lot of documentation.