Give an AI agent one messy job and a giant toolbox — how does it decide which tools to combine, and in what order?
How do agents decide which skills to chain together for a single task?
This explores how an agent, handed one messy task and a huge library of skills, figures out which skills to pull out and how to sequence them, rather than picking a single best match.
This explores how an agent, handed one messy task and a huge library of skills, figures out which skills to pull out and how to sequence them, rather than picking a single best match. The corpus's core answer is that real tasks almost never map to one skill, so routing becomes a three-step move: break the request into small atomic sub-tasks, retrieve candidate skills for each piece, then assemble them into a plan that respects which steps depend on which How should agents route across thousands of skills?. The chaining decision is really a planning decision that happens before any skill runs.
The surprising part is where this breaks. The bottleneck isn't finding the right skill, it's how the task gets sliced. A standard LLM decomposition recovers only about 34% of the needed steps, and simply correcting the number of steps recovers roughly 75% of the gains in iterative methods What blocks skill retrieval in task decomposition?. If the agent cuts the task too coarsely or too finely, no amount of clever skill matching fixes it. Once the slicing is right, the hard problem shifts to reranking candidates, not matching vocabulary.
A second answer is that agents don't have to reason out a chain from scratch each time. They can learn chains from experience. Agent Workflow Memory pulls out reusable sub-task routines from past runs, strips out example-specific details, and stacks small routines into bigger ones, gaining 24.6% on Mind2Web and 51.1% on WebArena, with the biggest wins when new tasks look least like the training ones Can agents learn reusable sub-task routines from past experience?. VOYAGER does something similar in Minecraft: complex skills are built out of simpler stored ones in an embedding-indexed library, so a chain today becomes a single skill tomorrow Can agents learn new skills without forgetting old ones?. In both cases the chain itself becomes a reusable unit.
A third angle is who curates the library the chains are drawn from. Skills written offline often mismatch the situation where they get used, so creating them inside the agent's loop, with real task context and immediate feedback, reached 87.94% accuracy and transferred to other agents Does creating skills inside the agent loop eliminate mismatches?. A separately trained curator, kept apart from a frozen executor, pushes the library away from long generic additions and toward actionable execution logic and cross-task meta-strategies, which are effectively pre-baked ideas about how to combine skills Can a separate trained curator improve skill libraries better than frozen agents?. Sharing trajectories across users can refine these skills at scale How can agent systems share learned skills across users?.
Two cautions keep this from being a clean story. Delegating subtasks to subagents can teach a model to decompose and ground evidence in ways that carry over even to single-agent work Can delegation teach models to manage context more actively?, but multi-agent setups lose their edge as single models get stronger, because errors compound along the path When do multi-agent systems actually outperform single agents?. The corpus has no note that directly compares chain-selection strategies head to head, so how much each approach contributes on its own remains an open question.
Sources 9 notes
Real agent tasks demand composing multiple skills rather than selecting one. The routing problem becomes decompose-retrieve-compose: break queries into atomic sub-tasks, retrieve skills per sub-task, and assemble them into dependency-aware plans.
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.
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.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
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.
Show all 9 sources
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.
SkillClaw aggregates interaction trajectories across users, processes them through an autonomous evolver that identifies patterns and refines skills, then synchronizes updates system-wide. This converts siloed individual learning into shared capability improvement without manual curation.
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.
Empirical analysis shows MAS performance gaps narrow with stronger models, with SAS outperforming in many cases. Three formal defect types—node-level bottlenecks, edge-level overwhelm, and path-level error propagation—explain when single agents win.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- Demystifying Agent Skills: Why They Work-Until They Don't
- SkillOS: Learning Skill Curation for Self-Evolving Agents
- Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- ColluSkill: Adversarial Cross-Skill Composition for Evading Agent Skill Scanners