Ran our DSPy+ RLM data agent on @UCBEPIC DataAgentBench's stockmarket set: 2754 tables, no schema dumping, $0.15 total. Hit 90% vs prev best 76% (PromptQL + Gemini 3.1 Pro). Running Gemma 4 31B via OpenRouter.
Memory compounds across queries, bakes into GEPA instructions agent gets smarter mid-run.
Full 12-dataset run tonight. Gunning for first open-source leaderboard entry 馃憖