If you use AI to boost your work today, does that actually make you better at it tomorrow — or just faster right now?
Which skills transfer across automatable and non-automatable work types?
This explores which skills carry over between work that AI can take over and work it can't. The corpus answers this from two directions: studies of human workers using AI, and research on how AI agents themselves learn skills that transfer.
This explores which skills carry over between work that AI can take over and work it can't. The collection has only a little direct evidence on human workers. What it has is not encouraging. When workers used generative AI on content tasks, they did much better in the moment. When they later did similar tasks on their own, they showed no improvement Does AI assistance help workers learn lasting skills?. The AI raised their output without raising their skill. At the level of the job market, job-vacancy data from ten countries show demand for AI skills clustering around a technical core: Python, SQL, machine learning and data analysis. Non-technical jobs are not moving toward that core. They are moving away from it Is AI creating common skills across jobs or deepening divisions?. So the skills employers ask for in AI-exposed work don't seem to be spreading into a shared base that every job draws on.
The research on AI agents learning skills offers a useful parallel. It suggests the skills that travel are about how to work, not what to know. A study of more than 8,000 agent trials found that skills mostly helped by giving the agent a steady procedure to follow (about 66% of cases). Supplying missing facts accounted for only about 5% Do skills teach procedures or inject missing facts?. A related finding is that instruction tuning mostly teaches models what a good answer looks like, not the task itself. Models trained on meaningless or even wrong instructions scored about the same as those trained on correct ones Does instruction tuning teach task understanding or output format?. Agent skills, at least, transfer as habits of structure rather than stored knowledge.
The skills that transfer most clearly sit one level up: breaking work into pieces and managing what you pay attention to. Models trained to hand off subtasks to helper agents and combine the results got better even on tasks with no helpers involved. The training seems to teach disciplined task-splitting and checking claims against evidence, not just coordination Can delegation teach models to manage context more actively?. Agents that pull reusable sub-routines out of past work gain more the more a new task differs from the ones they trained on Can agents learn reusable sub-task routines from past experience?. Getting the size of the pieces right turns out to be the main bottleneck. Standard task-splitting finds only about a third of the steps that matter What blocks skill retrieval in task decomposition?. Skill libraries that a trained curator maintains also drift away from task-specific tips toward strategies that work across tasks Can a separate trained curator improve skill libraries better than frozen agents?.
Put the two directions side by side and a pattern emerges. Doing a specific task, the part AI handles well, transfers poorly for both humans and models. Judgment that sits above any one task transfers well: how to split up a problem, when to hand work off, and how to check results. One more finding sharpens this. Skills built in the middle of real work, with immediate feedback, transfer better than ones written offline Does creating skills inside the agent loop eliminate mismatches?. That fits the worker study: if the AI does the task for you, you never build the skill. Treat this as an inference from parallels, though. The collection has no study that directly compares which human skills carry across automatable and non-automatable work.
Sources 9 notes
Wu et al. found that workers using generative AI performed substantially better on content tasks, but when performing similar tasks independently afterward, their performance showed no improvement. The capability did not transfer across contexts.
Vacancy data from ten countries show AI skill demand concentrating heavily within STEM occupations around Python, SQL, machine learning, and data analysis, while non-technical occupations diverge from this core rather than converge toward it.
Analysis of 8,135 trials shows procedural anchoring accounts for 65.7% of skill cases versus 4.5% for knowledge injection. Skills fail when retrieved incorrectly, invoked out of context, or followed too rigidly.
Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.
SearchSwarm shows that training models to delegate subtasks and integrate summarized results beats passive compression, with a 30B model matching much larger ones. Critically, the delegation skill transfers to single-agent tasks, suggesting it teaches disciplined decomposition and evidence grounding, not just orchestration.
Show all 9 sources
Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.
Standard LLM decomposition reaches only 34% step-level recall, gating retrieval success. Correcting step count recovers 75% of gains in iterative methods, shifting the bottleneck to representation-level reranking rather than vocabulary alignment.
SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.
MUSE-Autoskill demonstrates that invoking skill creation from within the agent's reasoning loop grounds new skills in exact task context, immediate feedback, and runtime validation. In-loop skills reach 87.94% task accuracy and transfer to other agents with minimal loss, eliminating the situated context problem of offline authoring.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Demystifying Agent Skills: Why They Work-Until They Don't
- RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- SkillOS: Learning Skill Curation for Self-Evolving Agents
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose
- Agent Workflow Memory
- How AI Impacts Skill Formation