Artificial Intelligence · arXiv:2611.01386 v1 · submitted 1 Nov 2026 · 28 references
Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces
The paper says
“Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions).”
more of the abstract
Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. We evaluate Zep Agent Skills on τ3-bench banking, AppWorld, and EnterpriseOps-Gym. Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. Most of the gain is on the tasks that the teacher solved. On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 4 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. Thorough production system for compiling agent skills from verified traces, with honest evaluation and clear limitations; the τ3 confound is known and does not undermine the overall pattern across benchmarks. This line is wrong
1 major objection · 7 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 28 references say.
The paper says, at §6.5, sentence b397
The review did not quote this passage.
Pith says: Objection 1 of 1 · major
The paper acknowledges that the τ³ Skills contain discoverable tool names and states that the experiment 'does not separate the effect of these names from the effect of the procedures.' This is a load-bearing confound for the τ³ results. Because the teacher solved tasks that require finding these names, the compiled Skills effectively supply them to the executor, which could explain a substantial part of the pass-rate gain (e.g., the +17.0 point gain on teacher-solved tasks in §7.5).
more of objection 1
The paper should attempt to quantify the contribution of name revelation, for example by comparing tasks with and without discoverable tools, or by conducting an ablation that removes tool names from the Skill bodies and re-running a subset of cells. If such an analysis is infeasible, the discussion should explicitly bound the magnitude of this confound and temper claims about procedural guidance for τ³.
7 minor objections
The paper says, at §5
Pith says: minor objection 1 of 7
In the second sentence of Section 5, 'codetrusted code' should be 'code—trusted code' (missing punctuation/spacing).
The paper says, at §6.2, paragraph 3
Pith says: minor objection 2 of 7
The paper mentions that the Skills arm differs from the no-Skills arm in three ways (Skill content, longer system prompt, extra tool) and does not separate these effects. The teacher strata analysis in §7.5 largely addresses this by showing gain is concentrated on tasks with a relevant Skill, but a brief acknowledgment of this reasoning in §6.2 would help readers.
show 5 more
The paper says, at Table 8 and Figure 4
Pith says: minor objection 3 of 7
In Table 8, four of eight cells have Holm-adjusted p-values > 0.05 despite all positive intervals. Consider adding a note that the pooled gain (+8.6 points, CI +5.5 to +11.9) is robust and that the individual p-values are conservative due to the small number of tasks (97) per cell.
The paper says, at §7.4
Pith says: minor objection 4 of 7
The latency analysis is explicitly confounded (acknowledged in §6.4 and §7.4). Consider moving the detailed latency discussion to an appendix or reducing its prominence, as the paper draws no conclusion from it.
The paper says, at §6.5
Pith says: minor objection 5 of 7
The τ³ Skill index contains 31 Skills. Consider adding a short table or list showing how many of those Skills contain discoverable tool names, to give readers a concrete sense of the scope of the confound.
The paper says, at Throughout
Pith says: minor objection 6 of 7
The notation for τ³-bench is inconsistent: the abstract and §6.3 use 'τ^3' while many figures and tables use 'τ3'. Unify to 'τ³' or 'τ3' for consistency.
The paper says, at References
Pith says: minor objection 7 of 7
Several citations (e.g., [Schmotz et al., 2026], [Liu et al., 2026b]) are arXiv preprints from 2026, which is unusual for a paper submitted in 2026. Verify that all citations are correctly dated and that no post-submission references are required.
What would settle it
Rerun the τ3 experiment with an additional control arm that provides only the discoverable tool names relevant to each task (a list of names matching the teacher's used tools) but no procedural steps. Compare the pass rate of this tool-names-only arm to the full Skills arm. If the tool-names-only gain accounts for more than half of the Skills gain, then the procedural content contributes less than the leaked names, and the claim that compiled procedures drive most of the gain is weakened.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 4 Nov 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee minor revision: 1 major objection, 7 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This paper describes Zep Agent Skills, a production service that compiles versioned, cited procedures from verified agent traces. The central claim—that compiled skills increase pass rate on repeated tasks—is supported by 26 experimental cells across three benchmarks, all showing positive gains, with appropriate statistical tests and confidence intervals. The paper is transparent about its limits: no transfer test in the main benchmarks, confounded latency analysis, and the acknowledged leak of discoverable tool names into τ3 skills. The stress-test concern about tool names on τ3 is real, but the authors flag it explicitly (sentence b397), and the pattern holds on AppWorld and EnterpriseOps-Gym where no such leak exists. So the overall claim stands.
What is new: The system architecture (separate Reflector and Curator roles, per-field citations, maturity tiers, search weight from verified use) is a thoughtful combination of known ideas with production-grade governance. The evaluation is more thorough than most system papers: multiple models, multiple reasoning efforts, paired comparisons, familywise error correction, and explicit leakage scans.
What the paper does well: Honest about limitations. Clear about threats (poisoning, leakage, cross-boundary flow) and the controls that address them. Releases skills in an open format. The ALFWorld comparison with RSD is illustrative rather than controlled, but useful context.
Soft spots: (1) The main evaluation is on replay of teacher-solved tasks. Transfer is only weakly evidenced (EnterpriseOps-Gym failed tasks, one AppWorld scenario, ALFWorld). The authors call this the biggest open question. (2) The τ3 tool-name confound is acknowledged but not quantified. (3) The benchmark harness uses a full Skill index, not production search, which inflates token cost on short tasks; the authors note this. (4) Compiler errors (merged opposite rules) cause some losses; the rate is not deeply analyzed. (5) The paper is an engineering report, not a new learning algorithm, so scientific novelty is moderate.
Who this is for: People deploying LLM agents who need practical procedural memory. Researchers in agent skill libraries will find the design decisions useful. A serious referee should engage.
Recommendation: Accept for peer review. Minor revision: quantify the tool-name confound on τ3 (even a rough estimate), and sharpen the statement about what the evaluation does and does not show. The transfer question remains open, but the paper sets clear expectations.
The strongest objection
The τ3-bench banking knowledge environment includes 'discoverable' tools whose names have numeric suffixes that the agent must find in the knowledge base. The compiled Skills contain these names (the paper acknowledges this in §6.5, sentence b397: 'The τ3 Skills contain the names of discoverable tools. The experiment does not separate the effect of these names from the effect of the procedures.'). The executor without Skills must spend steps querying the knowledge base to discover these tool names; with Skills, the names are supplied directly in the Skill body. This could be a significant (or dominant) source of the pass-rate gain on τ3. The teacher strata analysis (§7.5) does not control for this, because tasks the teacher solved are exactly the tasks whose Skills contain the correct tool names. If the gain on τ3 is largely from name revelation, then the central claim that compiled procedures increase pass rate is partly an artifact of information leakage rather than procedural guidance. The paper does not quantify this component. The concern is load-bearing because it directly affects the interpretation of the central claim: the pass-rate increase may be due to task-specific information (tool names) rather than generalizable procedural knowledge, which is the stated purpose of Skills.
The paper acknowledges the tool-name confound but does not control for it. This does not invalidate the overall finding—the full Skills system (names + procedures) clearly improves pass rates across all benchmarks. However, the central claim about 'procedures' compiled from traces is somewhat conflated with task-specific information leakage. The τ3 benchmark is the only one with such discoverable tools, and the gain on τ3 is a substantive part of the evidence (8 cells, 4 significant after correction). The concrete test would clarify whether the procedural content itself is necessary or whether simply revealing tool names suffices for a large part of the gain. Since the paper explicitly flags this limitation and the claim still holds for the full system (procedures plus names), the verdict remains ACCEPT; no adjustment is warranted. The reader's weakest assumption focused on compiler reliability, which overlaps only partially with this confound (compiler reliability affects whether names are correctly extracted; here the concern is about interpretability of the effect, not reliability per se). Hence partial agreement.
The referee's report
If the claimed results hold, the paper demonstrates a practical and effective approach to procedural memory for LLM agents that avoids the costs of fine-tuning and the information-retention limits of episodic memory. The system's design—particularly its emphasis on verified evidence, per-field citations, versioned admission with safety scans, and progressive-disclosure retrieval— advances the state of the art in agent skill lifecycle management. The evaluation is rigorous: paired comparisons across 26 cells (3 benchmarks × 3 models × multiple reasoning efforts) with appropriate statistical tests (bootstrap, Holm/Bonferroni correction), multiple executors, and explicit leakage controls. The ALFWorld comparison with RSD provides an informative baseline. The paper's transparency about limitations (no transfer test, tool-name confound, confounded latency) strengthens its credibility.
Pith's reading of the paper (a machine's summary, not the authors')
This paper describes Zep Agent Skills, a production service that automatically compiles, curates, and serves procedural memory for deployed LLM agents. The service records each task attempt as a Trace with ordered events, outcome, and verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills that cite their evidence, and admission gates control which candidates become active. On τ³-bench banking, AppWorld, and EnterpriseOps-Gym, Skills compiled from a teacher model's runs increase the pass rate of three weaker executor models at every reasoning effort by 4.5 to 18.8 points. On ALFWorld, Skills from a small model's own runs give a gain similar to or larger than continuous fine-tuning, with no training. The paper argues that verified procedural memory can let deployed agents learn from experience without retraining, making stronger models' successes reusable by cheaper models.
What carries the argument. The central mechanism is a two-role compiler: a Reflector (an LLM) writes proposals for new Skills or patches to existing Skills from verified Traces, and a Curator (another LLM) checks each claim against its evidence citations and current Skills. Trusted code validates structure, matches citations, runs scanners for instance data and instruction injection, and governs admission. Skills are stored as structured contracts with per-field citations and rendered into the open Agent Skills format for progressive-disclosure retrieval. This compile-curate-admit pipeline is designed to produce reusable, auditable, and safe procedures from verified task successes.
Core claim. The central claim is that a production service can autonomously compile structured, cited, and versioned Skills from strongly verified task Traces, and that these Skills increase the task pass rate of weaker executor models on the same tasks at every reasoning effort tested, with most of the gain coming from tasks the teacher solved. The claim is supported by 26 paired comparisons across three benchmarks, where the gain ranges from 4.5 to 18.8 percentage points, and by an ALFWorld experiment where Skills from a small model's runs remove 57% of failures from a weak base, outperforming the 37% removal reported for continuous fine-tuning. The paper thereby argues that verified procedural memory can serve as a practical alternative to fine-tuning for improving agent performance on recurring tasks.
The premise it rests on. The Reflector and Curator LLMs can reliably extract correct, generalizable procedures from verified Traces, and the scanners can detect instance data and instruction injections without false negatives, so that compiled Skills are safe and effective.
If the paper is right (Pith's inferences, not the authors')
Stronger models can solve a task once; weaker models can reuse the compiled Skill on later occurrences, potentially reducing cost and latency in production.
Skills increase not only overall pass rates but also reliability across repeated trials (pass4 and scenario pass rate improve in most cells).
On ALFWorld, Skills from a small model's own runs achieve a gain comparable to or better than continuous fine-tuning, without any weight updates, suggesting a lightweight alternative for self-improvement.
The teacher's success rate limits the gain: tasks the teacher fails gain little, so improving the teacher or learning from failed traces would broaden impact.
Long tasks see reduced tokens per passed episode when Skills are used, because the steps saved outweigh the fixed cost of the Skill index.
If Skills can be compiled from partially successful Traces, the gain on teacher-failed tasks may increase. The paper notes this as future work, but the current design depends entirely on verified successes. An inference: using sub-goal verification or segment-level success could extend the method to more tasks.
The paper's ALFWorld results suggest that self-compiled Skills could be used in a continuous learning loop without fine-tuning. An inference: a lightweight alternative to online weight updates exists, though the paper does not study repeated self-compilation cycles.
The loss example (merged opposite rules) indicates that merge conflicts are a risk for any compiler that combines evidence from multiple traces. An inference: better conflict detection or human-in-the-loop for ambiguous merges could reduce losses.
The benchmark setup uses a flat Skill index instead of search, so the token cost on short tasks may be an artifact. In production with search, the fixed cost would be lower, potentially making Skills beneficial on short tasks as well.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions). Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. We evaluate Zep Agent Skills on τ3-bench banking, AppWorld, and EnterpriseOps-Gym. Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. Most of the gain is on the tasks that the teacher solved. On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-03
The authors' main claim, as the classifier picked it out 0.99
“Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.75 · human participants 0.06 · controlled comparison 0.82 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.48
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions). | background 0.99 | 0.05 | 0.02 | asserted 1.00 |
| Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. | background 1.00 | 0.04 | 0.02 | asserted 1.00 |
| We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. | contribution 1.00 | 0.78 | 0.07 | asserted 1.00 |
| The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. | method 0.73 | 0.76 | 0.08 | asserted 1.00 |
| The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. | method 0.99 | 0.77 | 0.03 | asserted 1.00 |
| A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. | method 0.99 | 0.77 | 0.05 | asserted 1.00 |
| Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. | method 0.99 | 0.75 | 0.04 | asserted 1.00 |
| The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. | contribution 0.62 | 0.57 | 0.04 | asserted 1.00 |
| Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. | method 0.97 | 0.59 | 0.02 | asserted 1.00 |
| We evaluate Zep Agent Skills on $\tau^3$-bench banking, AppWorld, and EnterpriseOps-Gym. | method 0.59 | 0.94 | 0.04 | asserted 0.97 |
| Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. | result 1.00 | 0.94 | 0.57 | asserted 1.00 |
| Most of the gain is on the tasks that the teacher solved. | result 0.91 | 0.91 | 0.09 | asserted 1.00 |
| On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training. | result 0.99 | 0.89 | 0.18 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-03: 1 request, 9,472 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (9)
Figures are the authors', reproduced from the paper.
References (28)
- [4] URLhttps://arxiv.org/abs/2308.10144. Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, and Kunle Olukotun. Agentic Context Engineering: Evolving Contexts for Self... on PithPith's reviewarXiv
- [7] URLhttps://arxiv.org/abs/2309.02427. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as Operating Systems.arXiv preprint arXiv:2310.08560, on PithPith's reviewarXiv
- [8] URL https://arxiv.org/abs/ 2310.08560. Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A Temporal Knowledge Graph Architecture for Agent Memory.arXiv preprint arXiv:2501.13956, on PithPith's reviewarXiv
- [9] URL https://arxiv.org/abs/ 2501.13956. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language Agents with Verbal Reinforcement Learning.arXiv preprint arXiv:2303.11366, on PithPith's reviewarXiv
- [10] URL https: //arxiv.org/abs/2303.11366. Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Memp: Exploring Agent Procedural Memory.arXiv preprint arXiv:2508.06433, on PithPith's reviewarXiv
- [11] URL https://arxiv.org/abs/2508.06433. Dongge Han, Camille Couturier, Daniel Madrigal Diaz, Xuchao Zhang, Victor Rühle, and Saravan Rajmohan. LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation.arXiv preprint arXiv:2510.04851, on PithPith's reviewarXiv
- [12] URLhttps://arxiv.org/abs/2510.04851. Qirui Mi, Zhijian Ma, Mengyue Yang, Haoxuan Li, Yisen Wang, Haifeng Zhang, and Jun Wang. Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents.arXiv preprint arXiv:2602.01869, on PitharXiv
- [13] URLhttps://arxiv.org/abs/2602.01869. Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, Bo Zhang, and Liang He. AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution.arXiv preprint arXiv:2603.01145, on PithPith's reviewarXiv
- [14] URLhttps://arxiv.org/abs/2603.01145. Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Erchao Zhao, Xiaoxi Jiang, and Guanjun Jiang. Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills.arXiv preprint arXiv:2603.25158, on PitharXiv
- [15] URLhttps://arxiv.org/abs/2603.25158. Yubo Li. Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries.arXiv preprint arXiv:2607.10113, on PithPith's reviewarXiv
- [16] URLhttps://arxiv.org/abs/2607.10113. Qianjun Pan, Yutao Yang, Junsong Li, Jie Zhou, Kai Chen, Xin Li, Qin Chen, and Liang He. Anything2Skill: Compiling External Knowledge into Reusable Skills for Agents.arXiv preprint arXiv:2606.09316, on PithPith's reviewarXiv
- [17] URL https://arxiv.org/abs/2606.09316. Yuxuan Liu, Zhaochen Su, Lingyun Xie, Yuhao Zhang, Qing Zong, Jiahe Guo, Zhongwei Xie, Yiyan Ji, Yauwai Yim, Hongyu Luo, Xiyu Ren, Ruan Chenyu, Haoran Li, and Yangqiu Song. SkillRevise: Improving LLM- Authored Agent Skills via Trace-Condit... on PithPith's reviewarXiv
- [18] URLhttps://arxiv.org/abs/2609.15396. Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, and Tu Vu. WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution.arXiv preprint arXiv:2608.27454, on PithPith's reviewarXiv
- [19] URL https://arxiv.org/abs/2608.27454. Yuxing Lu, Yicheng Chen, Shanchan Wu, and Sercan Ö. Arık. Procedural Graphs: Self-Evolving Execution Structures for LLM Agents.arXiv preprint arXiv:2609.09153, on PithPith's reviewarXiv
- [20] URLhttps://arxiv.org/abs/2609.09153. 23 Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces A PREPRINT Model Context Protocol. SEP-2640 (draft): the io.modelcontextprotocol/skills extension. https://github.com/mod elcontextprotocol/modelcon... on PithPith's reviewarXiv
- [21] URL https://arxiv.org/abs/2610.27495. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains.arXiv preprint arXiv:2406.12045, on PithPith's reviewarXiv
- [23] URL https://arxiv.org/abs/ 2506.07982. Quan Shi, Alexandra Zytek, Pedram Razavi, Karthik Narasimhan, and Victor Barres. τ-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge.arXiv preprint arXiv:2603.04370, on PithPith's reviewarXiv
- [24] URL https: //arxiv.org/abs/2603.04370. Harsh Trivedi, Tushar Khot, Mareike Hartmann, Ruskin Manku, Vinty Dong, Edward Li, Shashank Gupta, Ashish Sabharwal, and Niranjan Balasubramanian. AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agent... on PitharXiv
- [25] URL https://arxiv.org/abs/2407.18901. Shiva Krishna Reddy Malay, Shravan Nayak, Jishnu Sethumadhavan Nair, Sagar Davasam, Aman Tiwari, Sath- wik Tejaswi Madhusudhan, Sridhar Krishna Nemala, Srinivas Sunkara, and Sai Rajeswar. EnterpriseOps-Gym: Environments and Evaluations for... on PithPith's reviewarXiv
- [26] URLhttps://arxiv.org/abs/2603.13594. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. ALFWorld: Aligning Text and Embodied Environments for Interactive Learning.arXiv preprint arXiv:2010.03768, on PitharXiv
show 8 more
- [28] URLhttps://arxiv.org/abs/2602.12670. Zhaorun Chen, Zhen Xiang, Chaowei Xiao, Dawn Song, and Bo Li. AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases.arXiv preprint arXiv:2407.12784, on PithPith's reviewarXiv
- [29] URL https://arxiv.org/ab s/2407.12784. Shen Dong, Shaochen Xu, Pengfei He, Yige Li, Jiliang Tang, Tianming Liu, Hui Liu, and Zhen Xiang. Memory Injection Attacks on LLM Agents via Query-Only Interaction.arXiv preprint arXiv:2503.03704, on PithPith's reviewarXiv
- [30] URL https://arxiv.org/abs/2503.03704. David Schmotz, Luca Beurer-Kellner, Sahar Abdelnabi, and Maksym Andriushchenko. Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks.arXiv preprint arXiv:2602.20156, on PitharXiv
- [31] URL https://arxiv.org/abs/2602 .20156. Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning, and Leo Yu Zhang. “Do Not Mention This to the User”: Detecting and Understanding Malicious Agent Skills in the Wild.arXiv preprint arXiv:2602.06547, 2026b. URLhttps... on PithPith's reviewarXiv
- [2020] URLhttps://arxiv.org/abs/2010.03768. Xiangyi Li, Yimin Liu, Wenbo Chen, Bingran You, Zonglin Di, Yifeng He, Shenghan Zheng, Kyoung Whan Choe, Jiankai Sun, Shuyi Wang, Chujun Tao, Binxu Li, Xuandong Zhao, Hejia Geng, Xiaojun Wu, Junwei Zhou, Xiaokun Chen, Hanwen Xing, Yubo Li, ... on PithPith's reviewarXiv
- [2023] URLhttps://arxiv.org/abs/2305.16291. Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent Workflow Memory.arXiv preprint arXiv:2409.07429, on PithPith's reviewarXiv
- [2024] URLhttps://arxiv.org/abs/2409.07429. Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. ExpeL: LLM Agents Are Experiential Learners.arXiv preprint arXiv:2308.10144, on PithPith's reviewarXiv
- [2025] URLhttps://arxiv.org/abs/2510.04618. Agent Skills. Agent skills specification. https://agentskills.io/specification , on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment