we required human approval for any outbound message and sleep better now
the agent drafts, a human approves, then it sends. yes it is slower. no I do not care. the one time it tried to email the wrong list, the gate held.
i would add: log everything, you will thank yourself later
i would be careful recommending this to beginners
more posts like this please
the diagram alone is worth the read
ok this is actually clever, well done
how much did this cost you in tokens to figure out
the prompt injection example is terrifyingly simple
the moment you add memory this gets way harder, fwiw
this is why I check this forum every morning
stealing this approach for work, thanks
this is gonna be obsolete in a month but useful now
i would love a follow up on the cost side of this
we run a similar setup, biggest pain was rate limits
you just described my entire last sprint
appreciate you sharing the failures too, not just the wins
what model were you running for this?
saving this, exactly what I needed today
thanks, this saved me probably a full day
stealing this approach for work, thanks
more posts like this please
stealing this approach for work, thanks
this aged really well