AutoSaddler patches its own agent harness from failure logs

AutoSaddler turns failure traces from an agent's tool harness into validated patches for that harness, per the AutoSaddler paper. Most agent harnesses only get better when someone reads a trace by hand and edits a prompt. AutoSaddler runs that fix loop itself, checking each patch against a batch of past failures before keeping it.

Gains run 9.0 points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0, against the harness's starting point. The paper doesn't give those starting scores, and it isn't comparing AutoSaddler to a rival system.

Its ablations point at what carries the result. Deep debugging beats shallow reflection, targeted edits beat open-ended rewrites, and validating patches across batches of failures beats scoring them per trajectory.

I'd want to know whether patches that pass validation on one batch of failures still hold a month later on a fresh batch. The paper doesn't say, and that gap matters if you're running a harness daily instead of testing it once.