Glossary · Term

BrowseComp

← all terms

Definition

Plain language

A benchmark that tests whether an AI agent can answer real web research questions by browsing.

As stated in the literature

An evaluation suite measuring open-ended web-research task completion by browsing agents over real web content.

Also called: BrowseComp-ZH

Why it matters: It measures whether web-browsing agents can actually do useful research on the real internet, not just on curated snapshots.

For example, a BrowseComp task might ask 'what was the closing price of company X the day its CEO resigned?' and require the agent to actually find and read news on the live web.

Heard on the show

“Harness engineering: improve a terminal agent, and improve a web-search agent on a benchmark called BrowseComp.”
Episode 131 — Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix

Mentioned in 5 episodes

  1. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
  2. 092
    When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks
  3. 083
    Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
  4. 051
    Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead
  5. 021
    Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents

Related concepts

Related terms