archive
· today in ai · 2026-07-14
What context does a coding agent need?
Archive item — written before sources were shown.
A SWE-bench study finds the signal for editing code lives in the code itself; summaries answered 4 of 45 behavioral questions versus 27 for raw source.
Holding code localization fixed with an oracle, the study varied only how surrounding code was represented and scored results against SWE-bench Verified. Natural-language summaries answered almost none of the behavioral questions the raw source answered (4/45 vs 27/45); compressed context matched whole files at a third of the tokens.
sources
- 01What Context Does a Coding Agent Actually Need to Act?arxiv.org · primary (paper)
Rami Steitieh