ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization

ActiveSaddler. Automated Curriculum Learning for Agent Harness Optimization. By Sungho Park and ten others, from POSTECH, KAIST, and one other institution. Published on the first of October. No code repository is linked to it.

An agent's harness is everything around the model: its prompts, its tool interfaces, its control logic. Automated harness optimization can substantially improve LLM agents by updating those pieces iteratively, using feedback from execution. Existing methods concentrate on how the harness gets updated. Which training scenarios produce the feedback behind those updates is largely held fixed. Yet as the harness evolves, the scenarios that are most useful for further optimization can change. A scenario that exposed a weakness early on may have little left to teach later. That suggests the training curriculum itself should adapt alongside the harness.

The authors frame that missing dimension as an automated curriculum learning problem, and introduce ActiveSaddler to address it. ActiveSaddler models the evolving curriculum as a non-stationary bandit: a repeated choice among options whose value shifts over time. Its optimization targets, the options it chooses among, are instantiated dynamically rather than fixed in advance.

Recurring failures are abstracted into reusable arms, one per failure pattern. For each pattern, the method estimates the learning progress that further targeting could bring. It then balances two pulls: revisiting weaknesses it already knows, and exploring unseen scenarios to find new ones. Every optimization outcome updates both the set of discovered failure patterns and their priorities, so the curriculum co-evolves with the harness it is shaping.

Experiments run on GAIA2 and on Terminal-Bench 2.0. The comparison keeps the harness optimizer the same and changes only the curriculum. Against that optimizer using a scenario order fixed before optimization, ActiveSaddler consistently discovers stronger harnesses, improving test Pass@1 by four point four percentage points on GAIA2 and by seven point five on Terminal-Bench.

Ablations then take the method apart. They show the gains depend on three things together: constructing optimization targets dynamically, estimating how their utility evolves, and balancing continued optimization against the discovery of new failures.

In the existing methods, the harness changed and the update rule was optimized while the scenarios largely stayed put. Feedback kept arriving from the same fixed sequence of situations, however much the agent had already learned from them. ActiveSaddler sets the scenarios moving too, choosing where to look next from what the last round of optimization revealed. The curriculum stops being a fixed input and becomes something optimized alongside the harness. On these results, the authors argue that automated curriculum learning is a crucial new dimension of harness optimization in its own right.