Do automated evals work? A comparison of LLM tools on app traces
HamelHusain · x · 2026-07-24
A discussion of whether automated evaluations actually work, centered on a method that uses multiple LLM-powered tools to find failures in application traces and compare them with domain-expert labels.
The attached chart compares several systems by recall, precision, discoveries, and false positives, showing the trade-offs between catching more issues and keeping noise low.
- Braintrust Loop: 87.2% recall, 79.1% precision, 20 discoveries, 9 false positives
- Arize AX Alyx: 74.4% recall, 91.0% precision, 19 discoveries, 3 false positives
- LangSmith (chat agent): 79.5% recall, 77.5% precision, 20 discoveries, 9 false positives
- Codex (GPT-5.5 High): 84.6% recall, 82.8% precision, 20 discoveries, 11 false positives
- Factory Droid (GPT-5.5 High): 84.6% recall, 83.3% precision, 17 discoveries, 10 false positives
- Claude Code (Claude Opus 4.8): 79.5% recall, 77.4% precision, 17 discoveries, 14 false positives
More from coding & agent
- Stanford's PaperFormalizer turns research papers into Lean-verified versions with author in the loop — StanfordUniversity · 2026-10-27
- Hugging Face TRL makes score centering + masked importance sampling default in async GRPO — _lewtun · 2026-10-27
- AutoResearch agent's weird training curve gets dug into instead of discarded — ZimingLiu11 · 2026-10-27
- Firecrawl offers 2.5x credits to a16z Speedrun founders, unveils $75M Series B — devdigest · 2026-10-27
- Chums: open-source Apple Watch app puts Claude Code and Codex on your wrist — brandon_galang · 2026-10-27
- Open-source Clairic Studio turns pasted text into typed data and working code in 5 languages — dfinke · 2026-10-27