AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces
Most agent reliability lives in the harness around the model, and AutoSaddler tunes that layer automatically while logging every change.
AutoSaddler is a systems paper that doubles as a governance story.
They show that a lot of agent reliability lives in the harness, the prompts, tools, and control logic around the model, and that you can tune it automatically from failure traces for about 9 to 10 points of gain.
The governance payoff comes for free. It keeps a full change log of every patch and why, a provenance record most agent stacks never keep.
AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces
AutoSaddler frames harness optimization as offline learning over the prompt, tool, and middleware layer, adds about 9 to 10 points of reliability with no model change, and logs every patch in an evolution DAG.