Microsoft researchers have released AutoSaddler, a system that improves AI agents without touching the model at all - it reads failure traces, writes patches to the scaffolding around the model, and keeps only the ones that hold up on validation. On three separate long-horizon benchmarks it improved the base harness by 9.0, 9.6, and 10.0 percentage points. The code is public.
Key facts
- Gains of +9.0 points on GAIA2, +9.6 on SWE-Bench Pro, and +10.0 on Terminal-Bench 2.0 over the corresponding base harnesses.
- The system changes no model weights - it edits prompts, tool configurations, and control logic, treating the harness as source code.
- Published August 24, 2026 by a 13-author team spanning Microsoft and academic collaborators, including Sungho Park, Jue Zhang, Qingwei Lin, Saravan Rajmohan and Dongmei Zhang.
- Primary source: AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces (arXiv:2608.23041); code at github.com/microsoft/AutoSaddler.
The paper opens with a problem anyone who has shipped an agent recognizes. As the authors put it, "LLM agents remain unreliable on long-horizon tasks, where small local failures can compound over extended interactions and lead to overall task failure." A single mis-parsed tool output at step 12 becomes a wrong assumption at step 40 and a failed task at step 130.
The known fix is a better harness - the layer of prompts, tool definitions, and control logic wrapped around the model. Harnesses demonstrably work. Ground Truth has covered how the harness, not the model, moved DeepSeek's score by twenty tasks. The trouble, in the authors' words, is that "harness design remains a manual and expensive process that requires searching over a large space of prompts, tool configurations, and control logic." It is skilled labor, it does not transfer between projects, and nobody enjoys it.
AutoSaddler reframes that search as an offline learning problem. Run the agent on a mini-batch of tasks. Collect the trajectories that failed. Diagnose why they failed, at the level of a specific decision the harness made or failed to prevent. Generate a structured patch - not a note to remember, an actual edit to the harness code. Then test the patched harness on held-out tasks and keep the patch only if it generalizes.
The analogy is close to a compiler with profile-guided optimization, or to a code review culture where every production incident produces a lint rule rather than a wiki page. What accumulates is not experience, it is enforcement.
The consistency is what makes the numbers credible. GAIA2 tests general assistant work with tools, SWE-Bench Pro tests repository-scale software engineering, and Terminal-Bench 2.0 tests command-line task completion. These are unrelated domains with unrelated failure modes, and one unchanged procedure lifted all three by roughly the same amount. A single tuned result on one benchmark would be noise. Three is a pattern.
The ablation study is where the actual finding lives, and the authors flag it as the takeaway: effective harness optimization "benefits from three ingredients: deep debugging rather than shallow reflection, targeted modifications rather than unconstrained editing, and generalization-aware selection rather than trajectory-specific repair."
Each of those is a rebuttal to something people are currently doing. "Deep debugging rather than shallow reflection" says that asking a model to reflect on what went wrong produces plausible-sounding diagnoses that do not fix anything; you have to actually trace the failure. "Targeted modifications rather than unconstrained editing" says that letting a model freely rewrite the harness makes it worse, because unconstrained edits break things that were working. And "generalization-aware selection rather than trajectory-specific repair" is the direct shot at memory-based approaches: patching for the specific run that failed teaches the harness that one run, not the class of failures it belongs to.
That third point deserves emphasis. There is a large and growing body of work on giving agents memory - stores of past episodes to consult before acting. AutoSaddler's result argues that the useful thing to persist is not the episode but the validated correction derived from it. One is a diary, the other is a rule.
The honest caveats are real. Optimizing a harness against benchmark validation sets is precisely the setting where overfitting hides, and "generalization-aware selection" is a claim about held-out performance judged by the same benchmark family it was tuned within. The paper does not report what happens on a benchmark the optimizer never saw, which is the question that matters for anyone deploying this.
There is also a well-established pattern in this exact corner of the field: large benchmark gains that thin out on contact with real work. Ground Truth has covered models that rewrite their own harness, gain 16 points, and flunk office work, and a new terminal benchmark that dropped the best agent from 84 percent to 34. Ten points on a benchmark is worth exactly as much as the benchmark is, and the code being public is the fastest way to find out which.
Originally published on Ground Truth, where every claim is checked against the primary source.
