A creeper broke GPT-6 Astra. It went to farm potatoes

GPT-6 Astra got further in Minecraft than any AI system before it. Then a single creeper wiped out its progress. Vals AI, an independent evaluation firm, streamed OpenAI's newest computer-use model on Twitch in a 141-hour run. Astra built a semi-automatic blaze farm and collected 6 Blaze Rods. It then found a Warped Forest, killed 6+ Endermen and came away with 3 Ender Pearls, the key items for reaching the End.

It put all of that in a chest. A creeper blew up the chest and Astra's bed, and everything was gone. Astra wrote itself a note: "ALWAYS CARRY CRITICAL ITEMS with keepInventory; don't store in unguarded chest ever again." Then it spent the next several hours farming potatoes. It was also visibly jumpy, logging "GREEN tall thing ahead was SUGARCANE, NOT creeper!" Viewers told it to pick up the pace. OpenAI's developer account reposted the thread.

This is funny, but it's also a real result. Astra was strong at planning a long task and weak at recovering from a setback. After one bad event, the agent switched to a low-risk loop that felt safe and did nothing for its goal. Anyone running agents for hours at a time will recognise that failure.

Why it matters: long-running agents break when a plan goes wrong, not only when the plan is bad, and a 141-hour Minecraft run shows that failure clearly.