tmgthb/Autonomous-Agents: A Chronological Index of LLM Autonomous Agent Research Papers

What the collection is and who it helps

Autonomous-Agents is a curated list of research papers on autonomous LLM agents, maintained by one author who adds new entries on a roughly daily cadence. The README describes it as autonomous agents research papers, updated daily, with a link to a separate Resources section.

The collection targets researchers who want to follow the field without building their own search pipeline. Academic search engines, arXiv, and conference proceedings each require a separate query and return results that mix relevant and irrelevant work. Autonomous-Agents provides a pre-filtered feed: papers that have already been identified as relevant to LLM-based autonomous agent research, listed chronologically with brief structured summaries.

The repository is organized as a reading list rather than a code project. There is no software to install, no API to call, and no output format beyond the Markdown files that make up the collection. The value it provides is editorial: the author has read enough of the field to identify what belongs in the collection and has committed to updating it consistently.

The MIT licence places no restrictions on how the list is used, cited, or incorporated into other work.

How the repository splits papers across batch files

The README for 2026 covers the fifth of five batches for that year. Earlier batches are linked at the top of the README as separate Markdown files in the resources/ directory. The full file tree of batch files covers: 2026 batches 1 through 5, 2025 batches 1 through 4, 2024, 2023, and an Earlier file for papers before 2023.

Each batch file is a sequential list of papers in chronological order. Papers within a batch are presented with the date of addition, a bolded title linking to the arXiv paper, and three bullet points: a one-sentence project description, a sentence on the methodology or architecture, and a sentence on the evaluation approach or key findings.

The resources/ directory at the top level holds the older batch files. The README itself holds the current 2026 batch 5. This structure means that a reader following the collection reads the README for current papers and navigates to resources/ for historical ones.

The split into batches prevents any single file from becoming unmanageably large, but it does mean that a search across the full collection requires either cloning the repository and using `grep` across all Markdown files, or reading each batch sequentially.

Navigating the collection: clone and search

The most efficient way to use Autonomous-Agents is to clone the repository and search the text locally. The plain-text Markdown format allows standard text search tools to work across all batch files at once. A researcher looking for papers on a specific topic, such as memory systems or multi-agent coordination, can search all files without loading each one individually.

For browsing, the GitHub interface renders each batch file as a readable page. The README links to all previous batch files at the top, making it possible to navigate the collection by year and batch from the repository's main page without cloning.

The Resources section, linked from the README, is a separate file in resources/ that the collection maintains alongside the paper batches. Its contents are not reproduced in the README, so a reader who wants both the paper list and the resources needs to navigate to that file separately.

The README source includes a BibTeX citation block for the collection under the key MaattaAutonomousAgents2023. It credits Teemu Maatta as author, gives the year as 2023, and includes a note field where the access date should be filled in. The README text itself does not render the BibTeX block, but it is present in the Markdown source for anyone who wants to cite the collection in a research paper.

What the 2026 batch 5 papers cover

The papers added in late June 2026, visible in the current README, cover several overlapping themes in autonomous agent research. Qwen-AgentWorld introduces a native language world model trained through a three-stage pipeline to simulate agentic environments across seven domains. MEMPROBE introduces a benchmark that evaluates LLM agent memory by reconstructing hidden user-state from a memory store, finding that task success is an insufficient proxy for memory quality.

MemClaw addresses governed shared memory in multi-agent systems, identifying four failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. A separate paper on Agent-Native Memory System Framework reviews 12 memory system architectures across five benchmark workloads, finding that no single architecture dominates all scenarios and that localized updates outperform global reorganization for cost-efficient maintenance.

OT-Agent presents a data curation pipeline for training agentic models through ablation of six stages. Data Recipes for Agentic Models studies the effect of trajectory length and task diversity on agentic benchmark performance.

These entries are representative of the collection's focus: papers that propose new architectures, benchmarks, or training methods specific to LLM-based agents operating over multiple steps or across sessions.

What the collection leaves out

Autonomous-Agents does not tag papers by sub-field, method type, benchmark, or model family. A researcher who needs all papers specifically on tool use in agents, or all papers evaluated on a specific benchmark, must do that filtering manually after downloading the collection.

The collection also does not include full abstracts, author affiliations, conference or journal information, or links to code repositories. Each entry provides three structured bullet points based on the abstract, not a full review or a link to the implementation. A reader who wants the paper's code needs to search for it separately on arXiv or GitHub.

Papers from top conferences such as NeurIPS, ICML, ICLR, and ACL are included when they are relevant, but the collection does not differentiate between workshop papers, preprints, and peer-reviewed work. All papers appear in the same list regardless of venue.

The collection is one person's view of relevance. Papers that fall outside the author's reading list will not appear even if they are significant. Treating it as a complete survey of the field would be a mistake; it is better understood as one well-maintained thread through a large literature.

Maintenance record and alternatives

The last push to Autonomous-Agents was on 2026-06-24. The repository has no GitHub releases and no versioned tags. New papers are added directly to the README or batch files as the author encounters them.

Papers With Code maintains a different kind of index: it tracks papers that release code and links each paper to its implementation. The difference is focus. Papers With Code covers all of machine learning and organizes by benchmark and task. Autonomous-Agents covers only autonomous agent research and organizes only by date. A reader who needs papers filtered by whether they release code, or who is interested in a broader ML topic, would find Papers With Code more useful. A reader who wants to follow the autonomous agent sub-field chronologically and without noise from unrelated ML topics will find Autonomous-Agents more directly focused.

The author provides a BibTeX entry for citing the collection and maintains a Twitter account linked from the README for updates.

Editorial conclusion

Autonomous-Agents is for researchers and practitioners who need a single chronological index of autonomous agent papers without sorting through arXiv searches or conference proceedings. It covers a specific domain well: LLM-based autonomous agent research from 2023 onward. It is not the right starting point for someone who needs papers on a narrow sub-topic, since the collection does not tag papers by sub-field or provide filtering. The last push was on 2026-06-24, and the repository has no formal releases. Anyone citing the collection in their own work should use the BibTeX entry provided in the README source and replace the YYYY-MM-DD placeholder with the date of access.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. tmgthb/Autonomous-Agents on GitHub