# PurCL/ASE tracks the agentic software engineering literature in one searchable page > The PurCL/ASE repository holds a curated database of 1,663 papers on agentic software engineering across software engineering, programming languages, security and natural language processing venues, with a two-level taxonomy, a static browse interface and a Claude Code skill that labels new proceedings automatically. **PurCL/ASE** — A continuously updated collection of papers on agentic SE - Repository: https://github.com/PurCL/ASE - Stars: 638 · Forks: 45 - Language: HTML - License: not declared - Published: 2026-10-21 · Updated: 2026-10-21 · Language: en - Canonical page: https://hysenlabs.com/projects/purcl-ase ## A literature database, updated by a pipeline PurCL/ASE is described as a curated literature database of 1,663 research papers on agentic software engineering, drawn from top-tier venues in software engineering, programming languages, security and natural language processing. The framing matters, because a paper list and a database are different things. A list is a set of citations someone collected. A database carries a schema, a classification and an interface for slicing it, and that is what this repository is. The second half of the description is the part with the engineering in it: the repository also provides an automated paper labeling skill, described as a Claude Code pipeline that extracts, filters and classifies new papers from raw proceedings files in various formats, for example bib and html, keeping the database up to date with minimal manual effort. The top level tree is consistent with that split, with a data directory holding the records, a web directory holding the browse interface, and a .claude directory holding the skill itself. The primary language of the repository is HTML, which tells you the interface is static and the data is a file, not a service. ## A static page with search, filters and AND logic You browse the collection by opening the index file in the web directory locally, or by using the hosted version. The interface supports full text search across titles and abstracts, independent single select dropdowns for year and venue, and a label filter that takes one or more research topics either from the sidebar or by clicking the label pills on a paper card. Every filter dimension is optional and they combine with AND logic, so only papers satisfying every active constraint remain visible. There are two details that make this pleasant to use. Venue names are normalised, so a single choice of ICSE matches every year of that venue, which removes the most tedious kind of filter UI. And each active constraint appears as a removable tag below the toolbar, so you can see at a glance what has narrowed the result set and take one piece of it back. Paper cards expand to reveal the abstract on click. There is no server, no account and no query language, and for a literature collection of this size that is a reasonable trade rather than a limitation. ## Coverage is systematic for recent years and best effort before that The collection rules are stated in one sentence and they set the boundaries of any claim you make with it. Papers are systematically collected for all proceedings from 2023 to 2026 that have been publicly released, and the database additionally includes selected papers from earlier years between 2020 and 2022, and other venues, on a best effort basis. That distinction is the one to keep in mind. A count of agentic papers at ICSE across 2023 to 2025 is a census of what that venue published in that period, while a number for 2021 is a sample, and a reader who treats the two as equivalent will draw a wrong conclusion about the trend. The tracked venue list is grouped by track. Software engineering covers ICSE, FSE, ASE and ISSTA, plus the journals TSE and TOSEM. Programming languages covers PLDI, OOPSLA, POPL, CC and COLM. Security covers S&P, USENIX Security, CCS, NDSS and RAID. Natural language processing covers ACL, EMNLP and NAACL, and machine learning venues ICML, NeurIPS and ICLR are included as well. The per venue year ranges are listed for each, and several of them stop at 2025, which is a fact to check before you describe a trend through 2026. ## The counts show where the field is crowded The README breaks the collection down by venue and year, and the shape of the numbers is the interesting part for anyone orienting themselves. In software engineering, ASE contributes the most with 292 papers across 2023 to 2025, ahead of ICSE at 166, FSE at 130 and ISSTA at 98, and that track totals 686. In natural language processing the totals are larger, with ACL at 294 and EMNLP at 250 for a track total of 566. Security accounts for 165 papers across 2023 to 2026, with NDSS the largest contributor at 55, and programming languages is thin at 41 with OOPSLA at 34. The imbalance is the interesting observation. Agentic software engineering papers appear far more often in the natural language processing venues than in the systems ones, which tells you something about how the field is publishing itself. The bars in the README are scaled within each track independently, with each block representing roughly nine papers, so the visual comparison is between venues inside a track and not across tracks. ## A two-level taxonomy with 47 sub-categories Classification is what turns a pile of papers into something you can query, and here the scheme is explicit. Papers are classified with a two-level taxonomy of nine top-level categories and 47 sub-categories, and a paper may carry multiple labels, so the counts below overlap by design. The taxonomy is grouped into three super-groups, the first of which covers papers where language models or agents are applied to core software engineering tasks. Code generation is the largest category with 750 papers, and its sub-categories include program synthesis at 410, program repair at 223, code translation at 69, code completion at 68, refactoring at 37 and decompilation at 23. Static analysis follows with 446, including bug detection at 249, code summarisation at 67, code search at 51, taint analysis at 16 and symbolic execution at 7. Dynamic analysis accounts for 287, code model for 427 with model training at 407, and other tasks such as log analysis, code review and commit message generation for 96. The second super-group covers agent architectures and the safety and security of code oriented models, with agent design at 328 split across planning, tool use, multi-agent and memory management, model safety and security at 267 covering adversarial attacks, jailbreaking, memorisation, backdoor detection and watermarking, and agent safety and security at 97 covering prompt injection, agent defence and access control. ## The labelling pipeline is the maintainable part A survey that has to be updated every few months is a survey that stops being updated, and the answer here is automation. The skill is described as a Claude Code pipeline, which means the procedure lives in the .claude directory in the repository and runs as a repeatable procedure rather than as a script someone has to remember to run. Its job is to take raw proceedings files, in formats such as bib and html, and extract, filter and classify the papers from them, writing the result into the data set. The web interface, the taxonomy and the automated labelling are the three parts of the design, and each one is a different kind of contribution: the interface makes the collection usable, the taxonomy makes it queryable, and the pipeline makes it current. For a project of this kind, that last piece is what determines whether the numbers in the tables above are a snapshot or a live figure. The README also has sections on adding new papers manually, contributing, extending the taxonomy, and a disclaimer and contact section, which together suggest a project that expects outside help with the parts a pipeline cannot do alone. ## Conclusion PurCL/ASE suits a researcher or a team lead who needs to understand where agentic software engineering work is actually concentrated, because the venue and category breakdowns turn a reading list into a map of the field, and the labels are specific enough, with separate counts for program synthesis, program repair, taint analysis, prompt injection and memorisation, to tell you which subfield is crowded and which is nearly empty. It is the wrong tool if you need a systematic review with a reproducible inclusion protocol, since the README describes the selection as systematic for recent years and best effort for earlier ones and for other venues. Verify three things before you cite it: the coverage window for the venue you care about, since several tracks are listed with year ranges that stop well short of 2026, how many papers sit in the categories relevant to you, given that the taxonomy is applied by an automated pipeline with human review, and the licence, since the repository listing reports no licence identifier. The last push was on 2026-07-17. ## FAQ ### What is the PurCL/ASE repository? It is a curated literature database of 1,663 research papers on agentic software engineering, drawn from top-tier venues in software engineering, programming languages, security and natural language processing, with a browse interface, a two-level taxonomy and an automated labelling pipeline. ### How current is the paper collection? Papers are systematically collected for all proceedings from 2023 to 2026 that have been publicly released, and the database also includes selected papers from 2020 to 2022 and from other venues on a best effort basis, so the earlier years are a sample rather than a census. ### How are the papers classified? With a two-level taxonomy of nine top-level categories and 47 sub-categories, organised into three super-groups, and a paper may carry multiple labels. The groups cover agents applied to core software engineering tasks, agent design and analysis, and the safety and security of code oriented models. ### How is the database kept up to date? The repository ships an automated paper labeling skill, described as a Claude Code pipeline that extracts, filters and classifies new papers from raw proceedings files in formats such as bib and html, so that the collection needs minimal manual effort to stay current. ## Sources - [Issues](https://github.com/PurCL/ASE/issues) - [PurCL/ASE on GitHub](https://github.com/PurCL/ASE) - [README](https://github.com/PurCL/ASE/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/purcl-ase