674 saved links, 2017–2026 (peak: 2024). This is your arxiv firehose — the tag where tweets pointing at papers land, 300+ of them arxiv.org and github.com alone. Three threads dominate: agent papers and surveys (your architect trajectory, footnoted), Anthropic’s research blog read like a magazine subscription, and a fast-growing genre of research about your own profession — what AI actually does to the people who build software. The 2024 peak is survey-heavy; the 2025–2026 saves get choosier and more agentic.
Related: LLMs · AI Agents · Deep Learning · Generative AI · AI (General)
The Anthropic shelf
You save Anthropic’s research posts the way other people save recipes — reliably, and across the whole range from interpretability to economics.
- Tracing the Thoughts of a Large Language Model — the interpretability landmark, and the introspection follow-up (“bruh even agents are introspective”).
- Building effective agents — the paper everyone building agents cites; you saved it early.
- How we built our multi-agent research system and effective harnesses for long-running agents — the engineering companions.
- Project Vend: can Claude run a small shop? — agentic AI meets a vending machine.
- The society thread: what 81,000 people’s economic hopes and worries look like · how people ask Claude for personal guidance · how agentic coding amplifies some knowledge and substitutes for other.
- Sleeper Agents — saved with the warning “this is not a hypothetical attack vector.”
Agents that do the research
The most self-referential thread in the vault: research systems that research. You’ve collected nearly every open deep-research implementation.
- LangChain’s Open Deep Research (video) and the repo — configurable planner/writer LLMs, search APIs, search depth.
- The Inventors of Deep Research — the Latent Space pod with the original Gemini PM and tech lead.
- Deep Research in Azure AI Foundry and building your own on Foundry — your Azure depth, applied.
- Sakana’s AI Scientist plus the paper explainer video — fully automated scientific discovery, 2024’s boldest claim.
- AI-Researcher: agentic scientific discovery · Paper2Agent — transforms research papers into agents · DeepResearchAgent.
- Open Scholar — Allen Institute’s open model that synthesizes science “as well as human experts.”
- The DIY tier: a fully local research assistant with Ollama · STORM over your own documents · Gemini + LangGraph fullstack research assistant · last30days-skill — researches any topic across Reddit, X, YouTube, HN.
- Rowboat — “persistent MD wikis are great for compiling research”; Karpathy liked it, and you’re literally living that thesis with this vault.
Research about your day job
The genre that grew fastest in 2025–2026: studies of what AI does to software engineers — career-relevant reading for someone rearchitecting their own role.
- Developer Productivity in the Age of Generative AI: A Psychological Perspective — “a colleague of mine wrote his master’s thesis on this,” says your note.
- The AI-Native Developer — which parts of software development are still worth doing?
- Anthropic’s experiment: coding with AI decreased skill formation and the counterpoint data on early-career developers — you’re collecting both sides.
- Hanselman & Russinovich: an apprenticeship model for junior engineers (ACM).
- DORA: balancing AI tensions — “perk up when the DORA team does research into software delivery.”
- AI in software engineering at Google: progress and the path ahead · what GitHub learned building Copilot.
- what matters when one engineer builds with agents: spec quality and domain knowledge, not AI smarts — saved via a Japanese thread; the finding travels.
- OpenAI’s own researchers: AI can’t solve most coding problems — the skeptic’s exhibit.
- Conducting Smarter Intelligences Than Me — the real insights from the Claude Code report, “100% human.”
Agentic AI, on paper
The academic backbone of your AI Agents tag — surveys, taxonomies, and the papers you flagged as foundational.
- The Hitchhiker’s Guide to Agentic AI — the 2026 book-length survey covering the whole landscape.
- Sutton & Silver: Welcome to the Era of Experience — “a must-read.”
- AI Agent ≠ Agentic AI — the Cornell paper on why the terms aren’t interchangeable; useful ammunition for architecture discussions.
- Surveys you kept: LLM agents: methodology, applications, challenges · evaluation of LLM-based agents · the visual-rich agent memory survey.
- SkillOpt: self-evolving agent skills — the skills thread, formalized.
- Google & MIT: a predictive framework for scaling multi-agent systems — there’s a tool-coordination trade-off, and you can optimize for it.
- Google’s whitepaper drip: the Agents whitepaper and Prototype to Production.
- Security: design patterns for securing agents against prompt injection (“a superb paper”) with Donato Capitella’s paper-review video.
- “The AI economy will be built on this paper” — your entire saved comment; bold claim, kept anyway.
Benchmarks, evals & judges
Your DevOps instincts showing: you don’t trust what you can’t measure, and this cluster proves it.
- Kaggle Benchmarks — “the potential to solve the biggest challenge in the LLM ecosystem: strong and diverse evals” — and Game Arena, models competing head-to-head at strategy games.
- Phil Schmid’s AI agent benchmark compendium plus the repo.
- Modal’s LLM Almanac — “we spent $100k benchmarking LLMs so you don’t have to.”
- FACTS Grounding — factuality benchmark · FrontierMath — the hard-math frontier · artificialanalysis.ai — speed and cost across API providers.
- The judge sub-thread: Awesome LLM Judges — the curated literature · Agent-as-a-Judge · Judge Arena — benchmarking the judges themselves.
- Microsoft’s open-source evals for agent interop and “unit tests” for AI agent skills — evals moving into CI, your home turf.
- Generative Benchmarking — Chroma on generating evals from your own data.
Why models do what they do
- OpenAI: why language models hallucinate with the five-revelations digest.
- On test-time compute and on long context — “two good papers landed today,” saved as a pair.
- Programming by Backprop — LLMs acquire reusable algorithmic abstractions during code training.
- Awesome Interpretability in LLMs — the field’s fast-moving reading list, curated.
- Deciphering language processing in the human brain through LLM representations — Google Research crossing into neuroscience.
Where research meets the racks
A quietly distinctive sub-thread: papers and lab posts about infrastructure — the corner of AI research where your two careers overlap.
- Regression language models predicting Borg performance from raw text — LLMs applied to Google’s own compute infrastructure.
- Google Cloud: preparing infrastructure for the agent-native era — what changes from cloud-native.
- awesome-LLM-AIOps — curated LLM research for IT operations.
- Datadog: don’t let LLMs write your post-mortems — the investigation is the learning exercise.
- Load balancing with random job arrivals and the evolution of graph learning since PageRank — the classic-systems reading you still make time for.
Learning to read the literature
The meta-skill: you’ve saved almost as much about how to consume research as research itself.
- The Ilya Sutskever → John Carmack reading list — “if you really learn all of these, you’ll know 90% of what matters today.”
- How To Read AI Research Papers Effectively (video) — the method.
- Sebastian Raschka’s noteworthy papers of 2024 and his instruction-pretraining roundup — the recurring digest you actually read.
- Pen and Paper Exercises in Machine Learning — the free 211-page PDF; you saved it twice, a year apart.
- ML/AI research papers — solved and From 0 to Research Scientist — the ramp-up repos.
- Foundations of Large Language Models — the free book: pretraining, scaling laws, RLHF · start with word2vec — the where-to-begin advice, Jeff Dean co-authoring.
- The Annotated Transformer — the canonical line-by-line implementation, still the best way in.
- Manuel Blum’s advice to a graduate student — “I like coming back to this and finding new insights every time.”
- Lessons on reviews and rebuttals — for when you’re on the other side of the paper.