Should an AI's skills be tailored to one job or built to work everywhere? Evidence favors pieces in between.
How do task-agnostic and task-oriented skills differ in coverage and reuse?
This explores how skills built for one specific job (task-oriented) differ from skills meant to work across many jobs (task-agnostic) in what they cover and how often they can be reused.
This explores how skills built for one specific job differ from skills meant to work across many jobs in what they cover and how reusable they are. The corpus has no note that compares the two head to head. Read together, though, its notes point the same way: the most reusable skills are neither whole-task recipes nor fully generic abilities, but mid-sized pieces. Agent Workflow Memory is the clearest case. It extracts routines at a finer grain than full tasks and strips out example-specific values so they can be reused elsewhere. That gave a 24.6% relative gain on Mind2Web and 51.1% on WebArena, and the gains grew as the gap between training and test tasks widened (Can agents learn reusable sub-task routines from past experience?). A recipe tied to one whole task only helps on that task. A sub-routine with the specifics removed helps on many.
The same split shows up inside a single reasoning system. When a decomposer (which breaks a problem into steps) is separated from a solver (which does each step), decomposition ability transfers across domains and solving ability does not (Does separating planning from execution improve reasoning accuracy?). So the task-agnostic skill is knowing how to carve up a problem, and the task-oriented skill is knowing how to do one particular kind of step. Being agnostic also doesn't mean training on one big general pool. Granite's function-calling model was trained on seven narrow subtasks, such as picking the function name, detecting parameters, and chaining calls, and it generalized better than models trained on umbrella datasets (Can breaking function calling into subtasks improve model generalization?). One caution: some of what looks like general task competence may only be knowledge of the answer format. Models tuned on empty or deliberately wrong instructions scored about the same as models tuned on correct ones (Does instruction tuning teach task understanding or output format?).
Coverage seems to come from combining specialists rather than from one broad generalist. Transformer2 keeps task-specific expert vectors and mixes them at inference. This covers new tasks without interference between experts and with fewer parameters than LoRA (Can models dynamically activate expert skills at inference time?). Skill documents work the same way at the text level. SkillOpt treats a skill document as trainable external state and accepts an edit only if it improves performance on held-out data. That check keeps a skill from being tuned to one task's quirks, and the resulting documents transfer across models (Can skill documents be optimized like neural network weights?). AIDE2 supports the general point: its gains held on four unseen benchmarks, including physics-based weather forecasting, which was outside the tasks it was selected on (Do AIDE2's improvements transfer to unseen tasks?).
Reuse has a cost, which is finding the right skill. A library of small, general pieces only helps if the agent can break a new task into steps that match those pieces. Standard LLM decomposition reaches only 34% step-level recall, and correcting the step count recovers about 75% of the gains in iterative methods (What blocks skill retrieval in task decomposition?). This is my inference from the notes rather than something they state, but it suggests the trade-off. A whole-task skill is easy to retrieve and rarely fits a new task. A fine-grained, agnostic skill fits many tasks but only if decomposition lands at the right step size. So the practical difference is less about which kind is better and more about how much work goes into slicing tasks correctly.
Sources 8 notes
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.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
Granite-20B-FunctionCalling shows that explicit training across seven granular subtasks—nested calls, chaining, parallel functions, name detection, parameter detection, next-best function, and response generation—generalizes better than umbrella datasets like ToolLLM. This multi-task approach closes the performance gap with GPT, Claude, and Gemini.
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.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
Show all 8 sources
SkillOpt treats skill documents as trainable external state of frozen agents, using a text-space optimizer with held-out validation gating to accept only edits that improve performance. Across 52 benchmark cells and seven models, the approach matches or exceeds baselines while adding zero inference cost and enabling transfer across models.
The paper reports that AIDE2's improvements transfer to four held-out benchmarks spanning machine learning, algorithm engineering, and physics-based weather forecasting—the last being outside the selection task distribution. This demonstrates transferable gains beyond overfitting to the selection set.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills
- Exploring Format Consistency for Instruction Tuning
- Divide-or-Conquer? Which Part Should You Distill Your LLM?
- Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose
- Demystifying Agent Skills: Why They Work-Until They Don't
- An Emulator for Fine-Tuning Large Language Models using Small Language Models
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- Distilling LLMs' Decomposition Abilities into Compact Language Models