Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-25 · high
The system diagnoses failure traces, rewrites harness code as a patch, then keeps only changes that hold up across a batch of new failures.
Why now: The paper posted to arXiv in August 2026, as more agent teams look for automated fixes to failures instead of hand-tuning the harness after every failure.
AutoSaddler treats an AI agent's harness, the code that manages its tools, prompts, and control flow, as something to patch automatically. When the agent fails a task, the system diagnoses the failure trace, writes a structured patch to the harness's own code, then validates the patch against a batch of new failures before keeping it. Patches that don't hold up get discarded.
Most agent harnesses get tuned by hand, one engineer fix at a time, after someone spots a pattern in the failures. AutoSaddler automates that loop, and the numbers in the AutoSaddler paper back it up: 9.0 points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0, all measured against the base harness before its changes.
The ablations explain why. Deep debugging beat shallow reflection. The system needed a real diagnosis of what broke, more than a quick self-critique. Targeted modifications beat unconstrained editing. Letting the agent rewrite whatever it wanted performed worse than constraining what it could touch. Generalization-aware selection beat trajectory-specific repair. A patch had to hold across a whole batch of new failures. Fixing only the one that triggered it wasn't enough.
The ablations matter more than the score gains. Unconstrained editing didn't help, so the improvement comes from limiting what the agent can change, not from giving it more freedom to rewrite its own code.
Each link below shares sources, entities, or timing with this story.
Evolved harnesses lost to a matched-budget sampling baseline once tested on a benchmark the search never saw.
HarnessRisk ran 128 sandboxed attacks across 14 model/harness setups and found configs that flagged the risk over 90% of the time still let it execute.
A trained failure monitor's accuracy collapses on a new model; a simple arithmetic check doesn't budge.
It treats harness improvement as offline learning: diagnose failure traces, generate structured patches that edit the harness as source code, then select updates by validation over mini-batches of failures (arXiv 2608.23041). Gains: 9.0 on GAIA2, 9.6 on SWE-Bench Pro, 10.0 on...
It automates the data-flow, crash-semantics, and commit-history work engineers do by hand.
A new checkpoint system saves an agent's memory and its environment together, so a failed attempt feeds the next try instead of getting discarded.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.