# Authors propose SkillCache for repeated LLM-agent skills [Canonical reading](https://escapemeridian.com/readings/authors-propose-skillcache-for-repeated-llm-agent-skills-d48585a8) Authors Jiayi Ni, Zac Wang, Alex Zhang, Zheng Liu and Roopa Prabhu ask whether an LLM agent can compile a reusable skill’s orchestration—the steps needed to carry it out—into a faster executable path. They report that compiling from instructions alone can produce brittle plans when assumptions about runtime bindings are wrong, making recovery costly. Their proposed SkillCache instead compiles a completed invocation into a parameterized capsule of programs and natural-language steps, replays it before reuse, and returns control to the model only when needed, keeping the original skill for recovery. In the Claude Code harness, warm execution of the same request on 87 SkillsBench tasks reduced tokens by 81% with Claude Sonnet 4.6 and 74% with Claude Opus 4.8 at comparable reward. The authors also report that capsules transferred across models: three open-weight models gained 12–25 percentage points in mean reward while using 72–81% fewer tokens. These results suggest a possible way to reduce the cost of repeated skill use and transfer skills across models, but the abstract’s reported evaluations do not establish how the approach performs beyond those settings. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 5, 2026 (2026-11-05). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-12T05:31:53.413Z. First observed: 2026-11-10T08:39:38.524Z. Retrieved: 2026-11-12T05:30:52.626Z. - [How Much Skill Orchestration Can LLM Agents Compile Away? Execution-Grounded Specialization for Token-Efficient Skill Use](https://arxiv.org/abs/2611.10022) ## Limits The evidence provided is the paper’s abstract, not its full methods or results; the reported findings are limited to the settings described there. The token reductions are reported for warm execution of the same request in the Claude Code harness on SkillsBench tasks, not as a general measure of performance across other settings. The abstract reports comparable reward in the Claude evaluations but does not specify the reward measure or provide the underlying results here. ## Open question How well would SkillCache work on requests and runtime conditions beyond those evaluated? ## Source-stated The authors say zero-shot skill specialization can make brittle plans when generated plans rely on incorrect assumptions about runtime bindings, with costly recovery. > incorrect assumptions about runtime bindings can make generated plans brittle and recovery costly - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The authors propose SkillCache, which compiles a completed invocation into a parameterized capsule and replays it before reuse, returning control to the model only when needed while retaining the source skill for recovery. > compiles a completed invocation into a parameterized capsule of programs and natural-language steps, replays it before reuse - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated In the Claude Code harness, warm execution of the same request on 87 SkillsBench tasks reduced tokens by 81% with Claude Sonnet 4.6 and 74% with Claude Opus 4.8 at comparable reward. > reduces tokens on 87 SkillsBench tasks by 81% with Claude Sonnet 4.6 and by 74% with Claude Opus 4.8 at comparable reward - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The authors report capsules transferred across three open-weight models, increasing their mean reward by 12–25 percentage points while reducing token use by 72–81%. > raising three open-weight models' mean reward by 12-25 percentage points while reducing their tokens by 72-81% - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The results suggest execution-grounded compilation may help with efficient repeated skill use and cross-model transfer. > suggesting that execution-grounded compilation is a promising approach to efficient repeated skill use and cross-model transfer - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source and contributor context Ni, Jiayi, Wang, Zac, Zhang, Alex, Liu, Zheng, Prabhu, Roopa · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Open weights A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted. - [Definition reference](https://opensource.org/ai/open-weights) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Editorial provenance Version 1. Publication ID: 32bbc755-3f28-41ea-8ba0-168e156f776c. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T05:31:53.413Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 8fb91c2c1d46458c7d071e2bc716522db7e9aeb4fb1264bb71b1172aac940e04. Source text hash: a35159d62986e6de3c7b9d89fb1e43c5c0555ae20dfbbd71930e71b4cc63882e. ## Revision and source-check history ### Version 1 — 2026-11-12T05:31:53.413Z Authors Jiayi Ni, Zac Wang, Alex Zhang, Zheng Liu and Roopa Prabhu ask whether an LLM agent can compile a reusable skill’s orchestration—the steps needed to carry it out—into a faster executable path. They report that compiling from instructions alone can produce brittle plans when assumptions about runtime bindings are wrong, making recovery costly. Their proposed SkillCache instead compiles a completed invocation into a parameterized capsule of programs and natural-language steps, replays it before reuse, and returns control to the model only when needed, keeping the original skill for recovery. In the Claude Code harness, warm execution of the same request on 87 SkillsBench tasks reduced tokens by 81% with Claude Sonnet 4.6 and 74% with Claude Opus 4.8 at comparable reward. The authors also report that capsules transferred across models: three open-weight models gained 12–25 percentage points in mean reward while using 72–81% fewer tokens. These results suggest a possible way to reduce the cost of repeated skill use and transfer skills across models, but the abstract’s reported evaluations do not establish how the approach performs beyond those settings. ## Original source arXiv cs.AI · Article Original source date: Nov 5, 2026 (2026-11-05). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-12T05:31:53.413Z. First observed: 2026-11-10T08:39:38.524Z. Retrieved: 2026-11-12T05:30:52.626Z. - [How Much Skill Orchestration Can LLM Agents Compile Away? Execution-Grounded Specialization for Token-Efficient Skill Use](https://arxiv.org/abs/2611.10022) ## Limits The evidence provided is the paper’s abstract, not its full methods or results; the reported findings are limited to the settings described there. The token reductions are reported for warm execution of the same request in the Claude Code harness on SkillsBench tasks, not as a general measure of performance across other settings. The abstract reports comparable reward in the Claude evaluations but does not specify the reward measure or provide the underlying results here. ## Open question How well would SkillCache work on requests and runtime conditions beyond those evaluated? ## Source-stated The authors say zero-shot skill specialization can make brittle plans when generated plans rely on incorrect assumptions about runtime bindings, with costly recovery. > incorrect assumptions about runtime bindings can make generated plans brittle and recovery costly - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The authors propose SkillCache, which compiles a completed invocation into a parameterized capsule and replays it before reuse, returning control to the model only when needed while retaining the source skill for recovery. > compiles a completed invocation into a parameterized capsule of programs and natural-language steps, replays it before reuse - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated In the Claude Code harness, warm execution of the same request on 87 SkillsBench tasks reduced tokens by 81% with Claude Sonnet 4.6 and 74% with Claude Opus 4.8 at comparable reward. > reduces tokens on 87 SkillsBench tasks by 81% with Claude Sonnet 4.6 and by 74% with Claude Opus 4.8 at comparable reward - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The authors report capsules transferred across three open-weight models, increasing their mean reward by 12–25 percentage points while reducing token use by 72–81%. > raising three open-weight models' mean reward by 12-25 percentage points while reducing their tokens by 72-81% - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source-stated The results suggest execution-grounded compilation may help with efficient repeated skill use and cross-model transfer. > suggesting that execution-grounded compilation is a promising approach to efficient repeated skill use and cross-model transfer - [Supporting source](https://arxiv.org/abs/2611.10022) ## Source and contributor context Ni, Jiayi, Wang, Zac, Zhang, Alex, Liu, Zheng, Prabhu, Roopa · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Open weights A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted. - [Definition reference](https://opensource.org/ai/open-weights) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Editorial provenance Version 1. Publication ID: 32bbc755-3f28-41ea-8ba0-168e156f776c. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T05:31:53.413Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 8fb91c2c1d46458c7d071e2bc716522db7e9aeb4fb1264bb71b1172aac940e04. Source text hash: a35159d62986e6de3c7b9d89fb1e43c5c0555ae20dfbbd71930e71b4cc63882e. ### passed — 2026-11-12T05:31:53.413Z — anthropic/claude-opus-5.5 Publication: 32bbc755-3f28-41ea-8ba0-168e156f776c attribution — pass: The reading attributes the question, SkillCache method and reported results to the five listed arXiv authors, matching the source. Claims are framed as what the authors report or propose, not as established fact. > Authors:Jiayi Ni, Zac Wang, Alex Zhang, Zheng Liu, Roopa Prabhu summary — pass: The title, summary and question faithfully convey the abstract: the brittleness of zero-shot specialization, SkillCache's one-trace compilation, the token and reward results, and cross-model transfer. Orchestration is explained in plain language. Terms such as warm execution, tokens and capsule stay exact for tooltip explanation. The omitted note that capsules stay effective under controlled request changes would add context, but its absence does not distort the meaning. The question is a genuine open issue. > capsules stay effective under controlled request changes date — pass: originalPublished is 2026-11-05. This matches the arXiv v1 submission date, not the later feed date of 2026-11-10. > \[v1\] Thu, 5 Nov 2026 00:01:07 UTC (664 KB) support — pass: Each claim matches an abstract excerpt, including the 81%/74% token reductions at comparable reward and the 12-25 point reward gains with 72-81% fewer tokens. The closing significance sentence is hedged and mirrors the authors' own 'promising approach' framing. The limitations correctly note that the evidence is the abstract and that the token results are scoped to warm execution of the same request. > suggesting that execution-grounded compilation is a promising approach to efficient repeated skill use and cross-model transfer media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct way to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/d48585a88be1c1af101f51c4091b1fc382d86cf7e5b4d929f2206b374741b837/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/authors-propose-skillcache-for-repeated-llm-agent-skills-d48585a8.json)