Researchers have built a system that teaches AI agents to pick their own practice problems while they are being trained.
Most tools for tuning AI agents work by rewriting the agent's instructions, tool access, and decision logic based on how it performs on a set of test scenarios. The catch is that those test scenarios usually stay fixed for the whole process, even as the agent's weaknesses change. A new method called ActiveSaddler fixes that by letting the training lineup shift as the agent improves. It treats each recurring type of mistake as a separate option to focus on, estimates how much more there is to gain from drilling that mistake further, and decides whether to keep working a known weak spot or go looking for new ones. On two benchmark suites, GAIA2 and Terminal-Bench 2.0, agents tuned this way scored 4.4 and 7.5 percentage points higher on first-attempt task completion than agents tuned with a training order set in advance.
Most effort in agent-tuning research has gone into rewriting what an agent is told to do, treating the problems used to catch its flaws as a fixed backdrop. This paper's point is that once an agent's weaknesses move, the scenarios needed to expose new ones should move too, or the tuning process starts chasing problems the agent has already solved. It's a structural fix rather than a smarter model, and a 7.5-point jump from reordering practice problems is a bigger swing than many new-model announcements deliver.
It's a benchmark result, not a shipped product, but it quietly undercuts the pitch behind "self-improving agent" tools that train against a static set of test cases and call it done.