Observed arrival · 2026-10-07
Searchlight: Does web search help coding agents?
An open benchmark tests how different search providers affect AI coding agents on research and coding tasks.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- People evaluating web search for coding agents
- Worth noticing
- Only 7% of agent searches were phrased as plain questions; the project reports better results when original sources appeared in search results.
Field notes
The task catalog spans recent research briefs, code fixes, and questions with well-documented answers. Searchlight includes both no-search and built-in-search controls, and says providers use their default settings. Its report also finds that memory performed similarly to the best search provider on documented-knowledge tasks, while using up to four times fewer tokens.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue