Kaggle and Eigent launch AgenticRL-Bench: agents graded on fixing 19 real agentic-RL incidents
MeganRisdal · x · 2026-11-21
Kaggle, with EigentAI, announced AgenticRL-Bench, evaluating whether coding agents can diagnose and fix real incidents across the agentic RL stack (Miles, slime, verl, SGLang, Megatron, Harbor).
- Built from 22 post-training RL tasks; the Kaggle leaderboard scores the 19 verifiable ones, run on Modal with 3 attempts each, and a task passes only if every hidden check passes
- Tasks mirror real agentic-RL engineering: Verl Upgrade Gate, Reward Determinism under Load, Trial Fault Attribution, SQL Partial Credit Exploits
- Positions Harbor as both the tool and the benchmark medium
The first benchmark focused on RL-stack incident response for agents.
More from coding & agent
- Vibe-coded site lets anyone independently verify the claimed Riemann proof with one Python command — sudoraohacker · 2026-11-21
- Dev builds FPS horror game in an afternoon with Fable 5.2 — TAbrodi · 2026-11-21
- Google's Stitch Loop agent autonomously splits goals into KPIs and builds fixes with Gemini — testingcatalog · 2026-11-21
- AFK Pilot adds Gemini 4 Neon support, warns of Antigravity impersonation bans — PawelHuryn · 2026-11-21
- Eric Elliott: agent skills favoring pure functions yield less error-prone code — ericelliott_ · 2026-11-21
- Protein binder competition test: can an LLM orchestrator out-reason human scientists? — madeofAjala · 2026-11-21