Gear-evolved harness beats hand-tuned agents on Kaggle Titanic: 79.9% vs 72.96%
Pale-Training-6944 · reddit · 2026-11-15
An experiment on Kaggle Titanic shows workflow matters more than model or time: Codex+Astra hit 72.96% in 60 minutes, explicit XGBoost guidance got 77.27% in 90-100 minutes, while Gear's GEPA loop evolved a DSH harness reaching 79.90% with 13-minute candidate runs. The evolved harness encodes a four-stage data-science workflow — data understanding/planning, focused experiments with error analysis, model selection under a fixed protocol, and delivery — making the workflow itself evaluable and evolvable.
More from coding & agent
- Developer ditches ChatGPT desktop for T3 Code after third try, citing frictionless multi-model switching — aniketapanjwani · 2026-11-15
- DeepMind's AlphaProof Code formally verifies Signal crypto with 6M lines of Lean in 4 weeks — dl_weekly · 2026-11-15
- 16x Microsoft MVP builds native PowerShell agent on Microsoft's Intelligent Terminal — dfinke · 2026-11-15
- Raindrop lets coding agents create accounts before humans claim them, with write-only keys — New-Tailor4071 · 2026-11-15
- Write the eval runner before the agent: evals-first architecture with Ports and mock adapters — hwchase17 · 2026-11-15
- How do you evaluate agents whose skill library keeps evolving between test runs? — tahaan · 2026-11-15