The "context screwup, not brain screwup" line is the whole thing.
I build with this daily, and almost every agent failure I end up chasing is the model getting the wrong files, not the model getting dumb.
The counterintuitive part is that adding more context feels safer, so your first instinct may be to dump everything in, which almost always makes it worse.
The real skill is curation. I spend more time deciding what to leave out of the context window than what to put in. Your "new-hire-desk" example nails exactly why.
The "context screwup, not brain screwup" line is the whole thing.
I build with this daily, and almost every agent failure I end up chasing is the model getting the wrong files, not the model getting dumb.
The counterintuitive part is that adding more context feels safer, so your first instinct may be to dump everything in, which almost always makes it worse.
The real skill is curation. I spend more time deciding what to leave out of the context window than what to put in. Your "new-hire-desk" example nails exactly why.
Yeah totally agree about the agent failure. Thanks so much for the comment!