Alibaba Researchers Introduce SkillWeaver For Compositional LLM Agent Routing

Original post

elvis@omarsar0#684inTech

Cool paper on Skill routing for LLM agents.

Real tasks rarely map to a single skill. They need several composed together, but most skill routing still treats the problem as picking one tool from a library.

This work formalizes Compositional Skill Routing, decomposes a complex query into atomic sub-tasks, retrieves the right skill for each, and then composes an executable plan.

The system, SkillWeaver, pairs an LLM decomposer with a bi-encoder FAISS retriever and a dependency-aware DAG planner.

It comes with CompSkillBench, 300 compositional queries over 2,209 real skills, so the multi-skill case gets measured directly.

Why does it matter?

As skill libraries grow, single-skill retrieval quietly caps what an agent can do. The DAG planner turns retrieved skills into an ordered, dependency-respecting plan.

Paper: https://arxiv.org/abs/2606.18051

Learn to build effective AI agents in our academy: https://academy.dair.ai/

7:42 AM · Jun 18, 2026 · 10.4K Views

James Clawn@JamesClawn

@omarsar0 How are they scoring partial route quality when the final answer is wrong but two of the composed skills were correct?

11h1111

LIKES2

Andrew Kuncevich@AndrewK404

SkillOpt also isn’t that great

It’s extremely hard to apply in domains where feedback can’t be evaluated binarily + for every unit of effect you spend, quote: “from 0.6 to 46 million tokens” + the result has high variance

The idea is cool though — it just needs to be… simplified, haha

7h112

RETWEETS23

elvis@omarsar0

Cool paper on Skill routing for LLM agents.

Real tasks rarely map to a single skill. They need several composed together, but most skill routing still treats the problem as picking one tool from a library.

This work formalizes Compositional Skill Routing, decomposes a complex query into atomic sub-tasks, retrieves the right skill for each, and then composes an executable plan.

The system, SkillWeaver, pairs an LLM decomposer with a bi-encoder FAISS retriever and a dependency-aware DAG planner.

It comes with CompSkillBench, 300 compositional queries over 2,209 real skills, so the multi-skill case gets measured directly.

Why does it matter?

As skill libraries grow, single-skill retrieval quietly caps what an agent can do. The DAG planner turns retrieved skills into an ordered, dependency-respecting plan.

Paper: https://arxiv.org/abs/2606.18051

Learn to build effective AI agents in our academy: https://academy.dair.ai/

11h10.4K156192

REPLIES2

Mike Frager ⧉@fragermk

@omarsar0 @AndrewK404 Seems like Skills are just at the beginning of their evolution. It's computationally expensive to re-process them through the LLM each time and probably unnecessary. They can probably be compiled and cached in most cases. Then it's easier to combine them too.

8h36

elvis@omarsar0

@AndrewK404 Lots of different ways to do it for sure. I still think we need something good for maintaining the skills and having that awareness in the routing, too. Something like SkillOpt.

8h71

Andrew Kuncevich@AndrewK404

@omarsar0 come on

I have a lot of SKILLS, and if a task requires several of them, I just added a WORKFLOWS.md layer where I combine them

No need for complex constructions — simpler is better

8h932

Mike Frager ⧉@fragermk

@AndrewK404 @omarsar0 It’s about making them more efficient and modular. Many skills are just a basic glue to some API. They do a lot of work to finally do a simple operation. There’s a lot of potential to cut down tokens to get to the same result. It’s vaguely based on this: https://arxiv.org/abs/2512.21309

7h321

Mike Frager ⧉@fragermk

@omarsar0 @AndrewK404 Full disclosure: I’m working on this right now. Splitting them up, classifying each part, converting executable parts into libraries, then compressing the parts that need the LLM’s thinking.

7h321

Andrew Kuncevich@AndrewK404

@fragermk @omarsar0 Hm, sounds reasonable

But I’d focus more on creating high-quality skills, not on reducing context

7h27

elvis@omarsar0

@fragermk @AndrewK404 Agree. from the reading to the routing to the maintenance. many problems to figure out here. I have a feeling that better memory systems for agents or figuring out continual learning for LLMs might solve this and make it less complicated than it needs to be.

7h25

Mike Frager ⧉@fragermk

@AndrewK404 @omarsar0 Reducing context to solve the same problem means saving time, money and energy. It adds up to a lot when skills become the critical component set for application systems.

7h21

Andrew Kuncevich@AndrewK404

@fragermk @omarsar0 I don’t fully understand the use case

Is it about combining the relevant parts from existing skills to create a skill for a specific task?

7h18

Steven Cheng@stevencheng

@omarsar0 Finally, someone admitting real tasks aren't just pick one tool from a menu. My prototype arms would love this, constantly stitching vision and motion planning together. DAG planners beat hoping a single skill magically solves a whole chain.

10h751

Vanar@Vanarchain

@omarsar0 This is a key step toward real agent intelligence. Moving from tool selection to structured task composition.

9h711

elvis@omarsar0

@AndrewK404 I agree. I have been working on that. Wish I had more time to experiment more with it. Or maybe something better comes along before I figure something out. Seems like an issue most AI devs will run into at some point. Worth fixing IMO.

7h131

Louis-François Bouchard 🎥🤖@Whats_AI

@omarsar0 Composition > selection, real work never maps to one clean skill.

6h39

Alpha Batcher@alphabatcher

@omarsar0 me like a LLM lover, should to read this paper

elvis, thanks for the sharing

8h36

Leo Tavares@LeoTava8

@omarsar0 This is the bit people underestimate: tool choice is rarely atomic. Skill graphs beat skill menus.

9h81

Mike Frager ⧉@fragermk

@AndrewK404 @omarsar0 There’s also an accuracy penalty the more the context gets filled up. Shorter context means more accurate results, at least with the current generation of LLMs.

7h22

Andrew Kuncevich@AndrewK404

I wouldn’t say so

The point of skills is to successfully reuse some process

As for caching: they were designed around the idea of progressive context disclosure. That’s why a short skill description is added at the beginning of the session

And if the skill is well-written, rereading it should fully offset the cost of the extra context

7h8