AI agents can learn skills from their own past runs, but why do those skills often break in a new setting?

This explores why skills distilled from an agent's own past runs (its trajectories) tend to work only in the setting where they were recorded. The corpus has no paper that measures this failure head-on, but several notes point at the same cause. A trajectory is a record of what happened in one place, so the specifics of that place come baked into it.

The most direct evidence is that trajectories only help when they match the situation. Why do trajectories matter more than individual examples for in-context learning? finds that in-context learning of sequential decisions needs full or partial trajectories from the same environment level, and isolated examples are not enough. A trajectory is useful because it shows how one specific world responds to actions. Carry it into a different world and the responses no longer line up. Transfer in that setup comes from having the right kind of trajectories on hand, not from one trajectory being universal.

The fix that shows up most often is to strip the specifics out. Can agents learn reusable sub-task routines from past experience? induces routines at a finer grain than whole tasks and abstracts away example-specific values, such as the particular page or item a run happened to touch. Its gains grow as the gap between training and test tasks widens, which is what you'd expect if raw trajectories were the thing that didn't transfer. Grain matters on the retrieval side too. What blocks skill retrieval in task decomposition? reports that standard LLM decomposition reaches only 34% step-level recall. A skill can exist and still go unused because the new task gets sliced differently from how the skill was cut.

Another fix is to stop using trajectories as the raw material. Can source code replace experience as skill raw material? extracts skills from GitHub source code and checks them by reconstructing the code without seeing it. It outperforms trajectory-derived and document-derived skills (42.90 to 47.90 macro-average across eight benchmarks). Code carries no memory of one agent's particular run. When skills are learned by a curator, the same pattern appears. In Can a separate trained curator improve skill libraries better than frozen agents?, the trained curator moves away from generic, wordy additions toward execution logic and cross-task meta-strategies, and those skills carry over to different executor models and domains.

Trajectories aren't inherently untransferable, though. When they train a general objective or behavior at scale, transfer does happen. Can language models learn to simulate agent environments? trains on over 10 million trajectories with next-state prediction and transfers across seven domains. Can delegation teach models to manage context more actively? finds that a delegation habit learned in multi-agent settings carries over to single-agent tasks. In both cases the model learned a general behavior, either predicting what happens next or breaking work into pieces. The failure comes from treating a single trajectory as the skill, when a trajectory is only evidence for one.