Meta's ML Agent Ran 1,097 Experiments on a Search Ranking Model
shoumikgoswami · reddit · 2026-11-19
Meta researchers built an ML-engineering agent that proposes code changes, launches training jobs, and evaluates results for a large-scale search ranker, with engineers still setting objectives and reviewing final models.
- 1,097 experiments completed; the best model lifted search volume by 0.23% in a large A/B test (researcher-reported, not independently evaluated)
- Each experiment takes hours to return feedback, so the system runs them in parallel
- The agent keeps a list of failed ideas to avoid re-proposing them
A rare case of an autonomous experimentation agent operating on an industrial-scale model.
More from coding & agent
- Executor Launches an 'Agent Cloud' Connecting Gmail, GitHub, Stripe and More in One Place — lucasmeijer · 2026-11-19
- DataTalks.Club launches free 5-week course on GPU-less generative media agents — Al_Grigor · 2026-11-19
- Claude Academy's free 67-lesson API course covers tools, evals, RAG and MCP — ZabihullahAtal · 2026-11-19
- Anthropic Offers Free Claude Courses With Badges, From MCP to Agent Skills — ZabihullahAtal · 2026-11-19
- Codex refuses to write security advisories for bugs it found itself, dev says — matthew_d_green · 2026-11-19
- Claude-powered guide maps 6 GTA 6 money routes as pre-orders top $1.5B — socialwithaayan · 2026-11-19