ArXiv agent papers, sorted by someone with opinions

What it does This repository is a curated reading list that tracks research on AI agents. The maintainers trawl arXiv and cherry-pick papers they deem novel, sorting them into granular categories like planning, memory, tool use, multi-agent systems, and real-world applications such as GUI or financial agents. It is a bibliography, not a codebase: every entry is a markdown link to an external paper, tagged as a survey, benchmark, or recommended read.

The interesting bit Most “awesome” lists try to be exhaustive; this one explicitly avoids comprehensiveness in favor of novelty. The taxonomy is unusually practical, splitting the field by capability (e.g., self-evolution, safety), architecture (single vs. multi-agent, agent-ops), and application domain (web agents, embodied agents, deep research). It reads less like a library index and more like a research newsletter that happens to live in a GitHub repo.

Key highlights

• Papers are grouped into capabilities (reasoning, perception, self-correction), architecture patterns (single-agent, multi-agent, agent-ops), and applied domains (software, enterprise, world simulation).
• Entries carry emoji tags: 🔥 for recommended reads, 📖 for surveys, and ⚖️ for benchmarks, making meta-analyses easy to spot.
• The maintainers favor “distinctively new approaches” over completeness, acting as a subjective but stated filter.
• Includes a section for tutorials and lectures, acknowledging that sometimes a good talk is worth more than a preprint.

Caveats

• The repo is purely a reading list: no code, no search interface, and no automated subscription feed beyond GitHub notifications.
• The README and repository description disagree on update cadence (weekly searches versus biweekly updates), so expect regular but not clockwork additions.
• Selection is subjective; the “particularly interesting” threshold is set by the maintainers without a public rubric.

Verdict Grab this if you are a researcher or engineer drowning in agent literature and want a human to pre-sort the noise. Skip it if you are hunting for source code, reproducible experiments, or an exhaustive systematic review.