← sources

Google used 29 AI agents to build a 50-task benchmark for causal reasoning

source: reddit (reddit:artificial)published 2026-11-14 07:30ingested 2026-11-15 02:34

retained text

Google Research says 29 agents spent 11.5 hours turning 13 public tables into a 50-task benchmark. The problems ask a model to decide whether the data can identify a causal effect and, where it can, estimate that effect against a known answer.

The step I'd borrow is the review. Google says separate reviewers and auditors found clues in the wording and shortcuts that let models get answers without doing the intended work. The team turned those problems into tests for future drafts.

Suppose an agent gets the right answer because every question with a certain phrase was generated from the same template. Would you trust a second agent to catch that, or would you keep a human-written test suite that the agents never see?

[Source: Google Research](https://www.research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/)

document fetched 2026-11-15 00:19 · canonical https://reddit.com/r/artificial/comments/1xvwdwz/google_used_29_ai_agents_to_build_a_50task