What's this about?
Agents remain unreliable on long tasks because small local failures compound. External harnesses help substantially, but designing them is manual work. AutoSaddler derives durable harness updates automatically from the agent's own execution traces. 50 upvotes.
What exactly is a "harness", and why does it need optimising?
The harness is the scaffolding around the model — which tools it can call, how it retries, when it stops, what it writes down. Most agent progress in the past year came from harness design rather than better models. The problem is that harnesses are hand-tuned by watching failures, which does not scale. This paper automates that loop: read the traces, find the recurring failure, patch the harness.
What's the community saying?
The word that matters is *durable* — an improvement written into the harness, so it persists rather than living in one session's context. It lands the same day as Recuris below, both attacking long-horizon reliability from the scaffolding side rather than the model side.