Compact-Bench shows LLM context compaction keeps at most 16% of the details coding agents need
Paired full-history versus summary-only replays on SWE-bench repairs. Keeping strings still fails tasks.
Read the piece#ground truth
One published Frontier piece on this subject — open it below.
1 article · Full archive · Topics
Read this piece
This tag currently points at a single story.
One brief a week.
New releases, then the constraint.