SWE-2 Agent Beats Super Monkey Ball One-Shot Using 41.9M Tokens Across 295 Messages
silasalberti · x · 2026-09-13
- An agent running SWE-2 reportedly solved Super Monkey Ball in a single one-shot run: it wrote the game code, played it, and QA'd it autonomously.
- Run stats: 295 messages and 41.9M total tokens — about 2.02M fresh input tokens, 39.6M cached input tokens, and 293K output tokens, highlighting how heavily long agent runs rely on cached context.
More from coding & agent
- Practical recipe: session embeddings + HDBSCAN + LLM-named clusters for chat topics — leland_mcinnes · 2026-09-13
- Chromium source ships an .agent/ folder with agent instructions and skills — HankYeomans · 2026-09-13
- AI FDEs Must Turn Production Failures into Evals, Release Gates, and Evidence — ryanorban · 2026-09-13
- Dev Open-Sources Two ComfyUI Workflows for Per-Slot Reference Control in Nano Banana Pro — Cute-Appointment6874 · 2026-09-13
- Three qualitative benchmarks for testing if AI agents can design systems — tianyin_xu · 2026-09-13
- Simulating famous thinkers beyond 'act like X': grounding personas in documented beliefs — haybreaker · 2026-09-13