If you abort an agent that looks like it's failing, how often are you wrong?

About one time in six, going by a 2026 real-time monitor built to make that exact call: it watches an agent’s steps as they happen, compares them against recorded successful runs, and can flag or block a run in about a millisecond per step instead of waiting for it to finish. The monitor, OnTrack, catches loop-shaped patterns, a stalled step, a repeated tool call, by aligning the live trajectory against reference runs instead of scoring it afterward. Tested on SWE-bench coding trajectories, it looked at only the first eight steps and still ranked trajectories headed for failure below the ones headed for success. Acting on that signal with an abort policy saved about 18% of the compute that would otherwise go to runs that failed anyway, but 83% of the runs it interrupted were actually heading to failure, meaning roughly 1 in 6 aborts killed a trajectory that would have succeeded. What drives the cost of evaluating an agent is the same per-event bill an early abort is trying to cut; catching a bad trajectory early only helps if you’re willing to pay for being wrong about it sometimes.