himanshu(@himanshustwts):Got a bunch of DMs asking, “Is my long-running task a long-horizon task?” I think tokenmaxxing has made people very delusional. A model generating tokens for lets say 3 hours has not necessarily solved a 3-hour task. Similarly, giving an agent million-token context window doesn’t give it long horizon capability. IMO, the time horizon should refer to the difficulty of the task as measured by average human completion time (AHT) I think *dependency depth* is even more useful mental model here where a task becomes genuinely LH when decisions made later depend on the meaningful chain of states produced by decisions made earlier.

2026.09.17 15:15

Got a bunch of DMs asking, “Is my long-running task a long-horizon task?” I think tokenmaxxing has made people very delusional. A model generating tokens for lets say 3 hours has not necessarily solved a 3-hour task. Similarly, giving an agent million-token context window doesn’t give it long horizon capability. IMO, the time horizon should refer to the difficulty of the task as measured by average human completion time (AHT) I think *dependency depth* is even more useful mental model here where a task becomes genuinely LH when decisions made later depend on the meaningful chain of states produced by decisions made earlier.

himanshu@himanshustwts

2026.09.16 23:50

i have the same thought if you get good enough intuition on building long-horizon RL tasks on particular domains the field is maybe fairly new for high schoolers and undergrads but you will be champ once you start looking into agent traces and being creative.