Google used 29 AI agents to build a 50-task causal reasoning benchmark in 11.5 hours
andrewaltair · reddit · 2026-11-14
Google Research used 29 agents via Antigravity's Teamwork feature to spend 11.5 hours converting 13 public tables into a 50-task causal reasoning benchmark. Tasks ask whether the data can identify a causal effect and, if so, estimate it against known answers.
The notable part is the review loop: separate reviewer and auditor agents flagged wording clues and template shortcuts that let models pass without real causal reasoning, feeding those flaws back as tests for future drafts. The author raises an open question: can a second agent reliably catch template leakage, or should you keep a human-written holdout suite agents never see?
More from coding & agent
- Money-saving trick: let a cheap model do the work, then ask it to summon a pricier model to review — vista8 · 2026-11-14
- Open-source personal agent The Machine drafts everything but never sends without your OK — markloperman · 2026-11-14
- Open-source Orinyx tests healthcare AI agents against FDA labels, no LLM in the verdict — Practical_Green5775 · 2026-11-14
- I tested 1,682 public MCP servers: 51% have tool inputs with no description — abhiramreddykalluru · 2026-11-14
- Solo AI system sold for $18,880 watches 1,300 execs and surfaces $50k–$200k deals — aryanXmahajan · 2026-11-14
- AAA game dev builds MAC, a multi-agent council where GPT, Claude and Codex review each other — Hirox-rd · 2026-11-14