OPAL: An AI Agent Rewrites Your Optimizer's Code Mid-Training
lmoroney · x · 2026-10-22
Xingjian Bai and collaborators released OPAL (OPtimizer with an Agent in the Loop), an AI agent that watches training dynamics and rewrites the optimizer's source code as a run progresses.
- Across every language and image setting tested without a tuned recipe, the authors report it matched or beat the best one-shot recipe
- It self-guarded against loss spikes and overfitting
- Available as a pip package you can point at your own training code
The author advises starting with a small run where you know what good looks like, and reviewing every change the agent makes before trusting it on larger jobs.
Related event: EPFL Team Releases OPAL, an Agent-in-the-Loop Reactive Training Optimizer(10 posts)→
More from coding & agent
- One-shot RL repro: training Qwen3.5-4B to paint watercolour flowers with a judge model — charles_irl · 2026-10-22
- Open-source Claude Duet alternative Open-Duet hits 5.9k stars on launch day — matchaman11 · 2026-10-22
- LGWM: A World Model That Verifies GUI Agent Transitions and Detects Hijacked Screens — Jiaming Zhang · 2026-10-22
- Tracing AI agents with OpenTelemetry: costs, P95 latency, and gaps in one pipeline — silksong_when · 2026-10-22
- New pattern: sync X bookmark folders to agents for intent-driven auto-pilot jobs — vincent_koc · 2026-10-22
- Towards AI builds writing benchmark: 4 LLM judges, 1000 human A/B tests on AI slop — AI Engineer · 2026-10-22