# Agentic Web is a reading list with 21 development papers and no code > The SAIL-Research-Lab/agentic-web repository collects papers about agents on the web under eight topic headings, carried by 18 authors across 8 institutions. It ships no package, no release and no example code, so it is worth reading as a map of the field rather than as something you install. **SAIL-Research-Lab/agentic-web** — Agentic Web: Weaving the Next Web with AI Agents. - Repository: https://github.com/SAIL-Research-Lab/agentic-web - Website: https://arxiv.org/pdf/2507.21206 - Stars: 479 · Forks: 43 - Language: Unknown - License: Apache-2.0 - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/sail-research-lab-agentic-web ## A reading list, with nothing to install The tracked tree holds three things at its top level: LICENSE, README.md and a docs folder. There is no package manifest, no build script and no release, so there is nothing to install and no version to pin. GitHub reports no primary language for the repository, which matches what the tree contains. The homepage field points at an arXiv PDF, 2507.21206, and no entry in the list links to that PDF, so the paper and the bibliography are kept side by side rather than merged. Apache-2.0 covers the whole thing, which is the part that matters in practice: the taxonomy of headings and the entry format can be reused as long as the licence travels with the copy. The repository has 479 stars and 43 forks, so the audience is people who bookmark a list, not people who depend on a package. Fork it if you want your own ordering of the eight buckets; the 43 forks suggest that is what the audience does. ## Eight buckets, and the one the index puts first The table of contents near the top of the README names eight anchors: Agentic Web Development, Information Retrieval, Recommendation, Agent Planning, Multi-Agent Learning, Safety and Security, Benchmark, and Citation. Seven of them are buckets for papers; the eighth is a citation block, and that is what makes the list citable rather than merely browsable. Every anchor is an in-page link, so nothing in the index points at a file inside the docs folder. Agentic Web Development comes first and is the longest section, which sets the emphasis for a newcomer: the list treats the construction of agents for the web as the primary thread, and retrieval, recommendation, planning, multi-agent learning and safety as the neighbouring problems that work on it. A reader looking for a planning paper has to pass the development bucket first, so treat the order as a reading sequence rather than a ranking. ## Three ways to add a paper, one address for opting out The README asks for additions in three ways: fork the repository, add the entry and merge it back; report an issue on the tracker; or email shangding.gu@berkeley.edu. The same address handles removal, since authors who would rather their paper not be listed are asked to contact it. The three routes are not equivalent, and the difference shows up in the history. A pull request leaves a reviewable commit with an author. An issue leaves a thread that may or may not turn into a commit. An email leaves no trace in the repository at all beyond the resulting edit. So the list can grow without any commit attributable to a named maintainer, which is worth remembering when you cite an entry and want to know who checked the details of it. ## Eighteen names, eight institutions, three project leads The header carries 18 author names, each with a superscript pointing at one of 8 institutions: Shanghai Jiao Tong University, the Hong Kong University of Science and Technology in Guangzhou, the University of Liverpool, the University of California at Berkeley, the Shanghai Innovation Institute, the University of California at Davis, Virginia Tech, and University College London. Three names carry a star and are marked as corresponding authors, that is project leads: Shangding Gu at Berkeley, Weinan Zhang across Shanghai Jiao Tong and the Shanghai Innovation Institute, and Jun Wang at University College London. Pieter Abbeel, Dawn Song and Costas Spanos also sit under Berkeley, alongside the two leads. Most of the rest of the roster is spread across Chinese, British and US campuses, and the contact address for both additions and removals is a Berkeley one. ## Twenty-one entries in the first bucket, weighted to 2025 The Agentic Web Development section lists 21 papers, and 12 of them are dated 2025. The older anchors sit lower down: WebArena, described as a realistic web environment for building autonomous agents; ToolLLM for letting large language models master more than 16000 real-world APIs; API-Bank as a benchmark for tool-augmented models; ReAct on pairing reasoning with acting; Voyager as an open-ended embodied agent; Toolformer on models teaching themselves to use tools; SWE-bench on whether models resolve real-world GitHub issues; and HuggingGPT. Macrec, a multi-agent collaboration framework for recommendation, is the single 2024 entry. What the composition shows is a tilt toward the planning and retrieval layer: Plan-and-act for long-horizon tasks, WebDancer for autonomous information seeking, MA-RAG and ArchRAG for retrieval, and a systematic examination of deep research agents. ## Four link hosts, and three links that end in a question mark Link targets across the list are not consistent. Most entries point at arxiv.org PDFs, but Macrec goes to dl.acm.org, ReAct goes to par.nsf.gov, and Toolformer and HuggingGPT go to proceedings.neurips.cc. Three arXiv links also carry a trailing question mark in the source: 2508.13787 for BetaWeb, 2505.20096 for MA-RAG and 2502.09891 for ArchRAG. That is a query character left behind inside a Markdown link, and it shows entries get pasted in faster than they get checked. The entry format is thin in other ways too: each line gives a title, an author list that often stops at et al., and a year, but no venue and no abstract. So the list tells you what to look for and leaves you to work out where the work actually appeared. ## Two figures, and an argument that the eras blur Two figures sit between the index and the first bucket. Figure 1 is titled Web Evolution: From Directories to Agents, which frames everything after it as one step in a longer shift rather than a new beginning. Figure 2 is a timeline of three major eras, and its caption argues against a clean division: the eras are not strictly separated, transitions occurred gradually, features of one era often coexisted with the next, and technologies and business models frequently overlapped during those transitions. A reading list built around those captions is taking a position, that agentic web is a continuation of the directory and search era instead of a break from it. The eight headings then do the opposite job, keeping retrieval, planning and safety apart so a reader can find one without reading the rest. The claim is made once, in a caption, and never argued again per topic. ## Last push 2026-07-18, no releases, no open issues The repository is not archived, and the last push to the main branch landed on 2026-07-18, a little under three months ago. Its own note describes the project as under active development and asks for constructive comments. Two counters next to that read zero: the releases page is empty and the issue tracker has nothing open. For a repository whose README invites readers to report a missing paper through an issue, that combination suggests additions arrive as pull requests or as direct email, or that the list changes quietly. The practical consequence stays small. A curated bibliography has no version to pin, so a citation to one entry should carry the date you read it, and anyone reusing the headings should copy what they need rather than depend on this particular list persisting unchanged. ## Conclusion Read this repository as a bibliography with a taxonomy attached, not as a component. It hands you a defensible way to slice agentic web research into eight buckets and a working set of 21 development papers, and reading it costs nothing. Two things to check before relying on it: entries carry a year and a partial author list but no venue, and three arXiv links have stray characters in them, so verify a link before you cite it. If you need an index with versions, venues and release notes behind it, this is not that. ## FAQ ### What is the agentic web, in this repository's terms? The README does not define the term in prose. It treats agentic web as a research area it is investigating, and sorts the work into eight headings that run from Agentic Web Development through Benchmark, ending with a Citation section. ### What is an example of an agentic browser? The list names no commercial browser. The closest entries are research systems: WebDancer, framed as an autonomous information seeking agency, and WebArena, a realistic web environment for building autonomous agents. ### Is ChatGPT an agentic AI? Nothing in the repository settles that. The entries are papers rather than product classifications, and they group work by technique: tool use, long-horizon planning, retrieval augmentation, multi-agent collaboration and benchmarks. ### Does agentic AI exist yet? The project treats the area as still forming. Its own note says the repository is under active development, and in the first bucket 12 of the 21 entries are dated 2025, with 2023 and 2024 papers still kept alongside them. ### What does the list say about agentic web search? One entry is titled From Web Search towards Agentic Deep Research: Incentivizing Search with Reasoning Agents, dated 2025. The repository has no search code or configuration of its own, so the topic appears there as reading rather than as a tool to call. ## Sources - [Issues](https://github.com/SAIL-Research-Lab/agentic-web/issues) - [License: Apache-2.0](https://github.com/SAIL-Research-Lab/agentic-web/blob/main/LICENSE) - [Project website](https://arxiv.org/pdf/2507.21206) - [README](https://github.com/SAIL-Research-Lab/agentic-web/blob/main/README.md) - [SAIL-Research-Lab/agentic-web on GitHub](https://github.com/SAIL-Research-Lab/agentic-web) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/sail-research-lab-agentic-web