The Reward Hackers

Temporal Agents: We gave agents the ability to rewind time. One primitive, branch(instruction, ago), lets an agent roll back to any earlier point in its own trajectory, swap in a new strategy, and keep going from there. The old timeline dies, the new one takes over, hard cut. Most agents burn their entire compute budget moving in a straight line, committing to decisions, accumulating errors, unable to course-correct without starting over. We broke that constraint. But rewinding isn't free (every branch costs steps, and the budget is fixed) so the real question is whether an agent can learn when to use it: a meta-policy that decides, in the moment, whether to keep pushing forward or burn steps rolling back to fix a mistake made six moves ago. Temporal control as learned compute allocation, not brute-force search.