Microsoft's Agent World Is INSANE. I Ran It. The Agent FAILED (0/3)

Microsoft Research's WildAWM gives agents fully working fake computers with a checker. I unpacked one hard task, ran the released 4B agent, and read the verifier myself: failed, 0 of 3. The reference solution passes 3 of 3 on the same checker. 0:00 The task: four hops and a checker 0:36 What WildAWM is 2:20 A task and its verifier 3:32 Why my laptop can't run it + picking a RunPod GPU 5:18 Unpacking the world 6:19 The agent setup (and what differs from the model card) 6:55 The run, step by step 9:30 The verdict: 0 of 3, and the control 10:49 Cost and what to try Sources: github.com/MSR-Orchard