what does human in the loop mean?
human in the loop means a person makes a decision that can change what the ai does next. the agent can gather the evidence and propose an action.
you can accept it, correct it or stop there. put that decision before a mistake would become someone else's problem.

in one line
a human is in the loop when their judgment helps decide the outcome of an ai task.
an example
dennis finds two recordings with the same concert title. in this example, an agent proposes removing the shorter one. dennis asks for the running times and the recording dates. the short one is a rehearsal he wants to keep. he rejects the deletion while both recordings still exist.
give the reviewer a real decision
a useful request shows exactly what a yes would change. for the recordings, show both filenames, the dates and the reason for choosing one. “tidy the archive?” gives dennis too little to decide.
put the decision with someone who knows why a rehearsal might matter. watching a progress bar gives them no chance to save it.
where to put the pause in dusk bot
choose ask for approval for a bot handling risky file operations. when an approval card arrives, use allow or deny. you can steer a working bot with another message.12
connected apps have their own permissions. for a review of records, give the bot read only access. it can propose a change in chat while you make the edit yourself.3
when personal review breaks down
keep the job with a person if a reviewer cannot judge the evidence. dusk bot is a poor fit for discarding original recordings on title alone. even a careful reviewer needs a way to listen to the doubtful pair.
questions
does reading the report afterward count as an approval?
a later review can improve the next run. if you need to prevent a deletion today, put your decision before the file disappears.
does a human have to do every step?
no. the agent can gather filenames and dates on its own. reserve your attention for choices where the same evidence could reasonably lead to different outcomes.
sources
drafted with ai help from the dusk bot docs and the sources listed on this page, then checked against them. if something here is wrong or out of date, write to [email protected].
what is prompt injection?
is it safe to let ai control your computer?
what is a computer-use agent?
what is an ai agent sandbox?
claude code for email: clean up one slice first
claude connectors: connect the job you actually need