[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-anthropic-coding-agents-social-sciences-research-ops-en":3,"article-related-anthropic-coding-agents-social-sciences-research-ops-en":30,"series-research-63b0f38d-55b8-4abb-9664-c0d602f4ea23":84},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"63b0f38d-55b8-4abb-9664-c0d602f4ea23","anthropic-coding-agents-social-sciences-research-ops-en","Anthropic’s survey turns coding agents into research ops","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fanthropic\">Anthropic\u003C\u002Fa>’s survey shows where coding agents fit in social science research and how to use them.\u003C\u002Fp>\u003Cp>I've been using AI in research workflows long enough to know when something feels wrong. Chatbots were fine for polishing prose and kicking around ideas, but when I handed them a real analysis task, they’d nod along like an overeager intern. They’d draft code, sure. Then I’d spend half an hour checking the imports, fixing the file paths, and untangling whatever nonsense they’d quietly introduced. Helpful? Sometimes. Trustworthy? Not enough.\u003C\u002Fp>\u003Cp>What’s changed with coding agents is not that they write better sentences. It’s that they can actually sit inside the workflow and do work: read a dataset, write analysis code, run it, inspect the failure, retry, and keep going without me babysitting every step. That’s the part that got my attention when I read Anthropic’s post, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fcoding-agents-social-sciences\">Coding agents in the social sciences\u003C\u002Fa>. It’s not a hype piece. It’s a survey of 1,260 quantitative social scientists, and the numbers are annoyingly useful.\u003C\u002Fp>\u003Cp>And the numbers are the story. 81% have tried genAI for research, but only 20% regularly use coding agents. That gap tells me the field is still in the awkward phase where everyone has heard the pitch, but most people haven’t changed how they work. That matters, because the researchers who do adopt these tools are not using them for fluffy drafting. They’re using them to code, edit, and push work forward faster.\u003C\u002Fp>\u003Ch2>1) The real split is not “AI users” versus “non-users”\u003C\u002Fh2>\u003Cblockquote>“81% of respondents said yes. But ... only 20% of respondents use coding agents.”\u003C\u002Fblockquote>\u003Cp>What this actually means is that we need to stop lumping every AI tool into one bucket. A chatbot that helps me rewrite a paragraph is not the same thing as a command-line \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> that can write and execute analysis code. Anthropic is drawing a line between general genAI use and actual coding-agent adoption, and that line matters.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780315413286-4dxu.png\" alt=\"Anthropic’s survey turns coding agents into research ops\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The post says the survey was fielded in February and March 2026, with 1,260 quantitative social scientists. That’s enough to show a pattern, not enough to pretend it’s the whole profession. Still, the pattern is hard to ignore: people have tried AI, but only a minority have let it into the center of the research process.\u003C\u002Fp>\u003Cp>I’ve seen this in my own work. Teams say they “use AI,” but what they really mean is they ask for a summary or a code snippet once in a while. That’s not adoption. That’s dabbling. Coding agents are different because they can sit in the loop with the dataset and the analysis environment. That’s where the risk and the payoff both get bigger.\u003C\u002Fp>\u003Cp>How to apply it: split your own workflow into categories. I use three:\u003C\u002Fp>\u003Cul>\u003Cli>idea support: brainstorming, outlining, literature triage\u003C\u002Fli>\u003Cli>draft support: prose cleanup, explanation, formatting\u003C\u002Fli>\u003Cli>execution support: generating, running, and revising analysis code\u003C\u002Fli>\u003C\u002Ful>\u003Cp>If your tool never touches execution, don’t pretend it’s changing your research pipeline. It’s a helper, not an operator.\u003C\u002Fp>\u003Ch2>2) Coding agents are getting used where the work is, not where the marketing is\u003C\u002Fh2>\u003Cblockquote>“The most common use, for both coding agent users and others, is for coding up analysis of quantitative data: 97% of coding agent users and 77% of other AI users report using it to generate code.”\u003C\u002Fblockquote>\u003Cp>This is the part I trust most in the whole post. Everyone loves to talk about AI writing papers. Anthropic’s data says the real center of gravity is code. Editing prose comes next. Methods advice and background research follow. Drafting prose is not the dominant use case, and that’s a relief, honestly, because the internet has enough synthetic introductions already.\u003C\u002Fp>\u003Cp>What this actually means is that researchers are using AI where the friction is highest. Analysis code is tedious, repetitive, and easy to get slightly wrong. That makes it a natural fit for an agent that can iterate. If I ask a model to write a regression script, I don’t just want the first draft. I want it to run, fail, inspect the error, and fix the thing without me having to play traffic cop.\u003C\u002Fp>\u003Cp>I ran into this while prototyping a survey analysis pipeline. A chatbot gave me a decent start, but once the dataset got messy, the model started hallucinating variable names and inventing cleaning steps that were not in the data. A coding agent, by contrast, is more useful when it can see the files, inspect outputs, and revise the code based on actual runtime feedback.\u003C\u002Fp>\u003Cp>Anthropic’s mention of tools like \u003Ca href=\"https:\u002F\u002Fclaude.ai\u002Fcode\">Claude Code\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fopenai.com\u002Fcodex\u002F\">Codex\u003C\u002Fa>, and \u003Ca href=\"https:\u002F\u002Fcursor.com\u002F\">Cursor\u003C\u002Fa> is the right mental model here. These are not writing toys. They are workflow tools. If you’re using them like a smarter autocomplete, you’re leaving most of the value on the table.\u003C\u002Fp>\u003Cp>How to apply it: use coding agents first on tasks with clear inputs and measurable outputs:\u003C\u002Fp>\u003Cul>\u003Cli>load and inspect a dataset\u003C\u002Fli>\u003Cli>write one analysis script\u003C\u002Fli>\u003Cli>run tests or sanity checks\u003C\u002Fli>\u003Cli>summarize what changed and why\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That sequence is boring, which is exactly why it works.\u003C\u002Fp>\u003Ch2>3) Adoption is uneven for reasons that should make us uncomfortable\u003C\u002Fh2>\u003Cblockquote>“Those with typically male names have adopted coding agents at more than twice the rate of respondents with typically female names.”\u003C\u002Fblockquote>\u003Cp>Anthropic reports that researchers with typically male names use coding agents at more than twice the rate of those with typically female names. It also says researchers at top universities are 40% more likely than others to use them. Economists and political scientists are much higher adopters than public health, education, or communications. Early-career researchers are using them more than tenured professors.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780315404010-kwg9.png\" alt=\"Anthropic’s survey turns coding agents into research ops\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>What this actually means is that adoption is not just about interest. It’s about status, training, infrastructure, and maybe a little confidence theater too. If you already work in a code-heavy discipline, you have fewer barriers. If your department expects you to ship analyses fast, you have more incentive. If you’re junior, you have more pressure to produce. If you’re at a top university, you’re more likely to have access, peers, and permission.\u003C\u002Fp>\u003Cp>I don’t like pretending this is neutral. Tools always spread unevenly at first. But when the tool changes the shape of research labor, the unevenness matters more. If only some groups get faster at project setup, data analysis, and iteration, then the productivity gap compounds.\u003C\u002Fp>\u003Cp>The post is careful here: these are descriptive differences, not causal claims. Good. That caution matters. Users of coding agents may already be the people who publish more, apply for more grants, or feel more comfortable experimenting with new tooling. The tool may be amplifying existing behavior rather than creating it from scratch.\u003C\u002Fp>\u003Cp>How to apply it: if you’re leading a lab or team, don’t assume adoption will happen evenly on its own. I’d do three things:\u003C\u002Fp>\u003Cul>\u003Cli>pair a high-confidence user with a skeptical one on the same task\u003C\u002Fli>\u003Cli>document the exact workflow, not just the tool name\u003C\u002Fli>\u003Cli>track whether the tool reduces setup time, not just whether people “liked it”\u003C\u002Fli>\u003C\u002Ful>\u003Cp>If you want fair adoption, you have to design for it. Otherwise the people already ahead just get another gear.\u003C\u002Fp>\u003Ch2>4) Coding agents seem to help the front of the pipeline more than the finish line\u003C\u002Fh2>\u003Cblockquote>“Coding agent users are starting more projects, posting more working papers, submitting more grants, and possibly sending out more conference submissions.”\u003C\u002Fblockquote>\u003Cp>This is where the post gets interesting for anyone who actually ships research. Anthropic says coding-agent users look more productive in the early pipeline: they start more projects and post more working papers. But the same pattern does not show up clearly at the journal-submission stage. No evidence of more new paper submissions or faster resubmissions.\u003C\u002Fp>\u003Cp>What this actually means is that coding agents may be best at getting work moving, not necessarily at polishing it to the point where journals will take it. That fits my experience. The annoying part of research is often not the first draft of the analysis. It’s the accumulation of tiny decisions that turn an idea into a runnable project. Folder structure. Reproducible code. Cleaning. Re-running after a data update. Those are exactly the places where an agent can shave off time.\u003C\u002Fp>\u003Cp>But the last mile is different. Submission quality still depends on judgment, framing, robustness checks, and a human sense of what is defensible. I don’t see an agent replacing that. I do see it making the middle of the process less painful.\u003C\u002Fp>\u003Cp>Anthropic also notes that these are self-reported outcomes over the six months before the survey. That’s useful, but it’s not proof that the tools caused the difference. Still, even a descriptive signal matters if you’re deciding where to spend your own attention.\u003C\u002Fp>\u003Cp>How to apply it: use agents to accelerate the messy middle, not to declare victory at the end. I’d delegate:\u003C\u002Fp>\u003Cul>\u003Cli>project scaffolding\u003C\u002Fli>\u003Cli>analysis script generation\u003C\u002Fli>\u003Cli>result table assembly\u003C\u002Fli>\u003Cli>robustness-check variants\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Then I’d keep human control over interpretation, claims, and submission packaging. That split is practical, and it keeps you from outsourcing the wrong things.\u003C\u002Fp>\u003Ch2>5) The risk is not just bad output, it’s a new kind of research bottleneck\u003C\u002Fh2>\u003Cblockquote>“These tools could accelerate science and make it more daring... They could also amplify disparities in research resources and exacerbate congestion in the scholarly record.”\u003C\u002Fblockquote>\u003Cp>This is the part people skip because it’s less fun than the productivity story. Anthropic is saying the upside is faster discovery, but the downside is more inequality and more junk in the pipeline. I think that’s right.\u003C\u002Fp>\u003Cp>What this actually means is that if coding agents lower the cost of producing a paper, they also lower the cost of producing mediocre papers. That matters because research systems already struggle with overload. Faster drafting is not the same thing as better science. If anything, it can make weak work cheaper to produce.\u003C\u002Fp>\u003Cp>There’s another wrinkle here. If AI systems begin making more of the analytical choices, then those choices shape the field. Which defaults get used? Which cleaning steps get omitted? Which robustness checks get treated as standard? Those are not trivial implementation details. They become part of the epistemic machinery of the discipline.\u003C\u002Fp>\u003Cp>I’ve watched teams adopt a new tool and immediately standardize around whatever the model happened to suggest first. That is convenient and dangerous at the same time. Convenience is how defaults become doctrine.\u003C\u002Fp>\u003Cp>How to apply it: treat agent output as a draft of process, not just a draft of results. I’d require:\u003C\u002Fp>\u003Cul>\u003Cli>a written log of agent-generated steps\u003C\u002Fli>\u003Cli>a human explanation for every major analytic choice\u003C\u002Fli>\u003Cli>one independent rerun without the agent before anything ships\u003C\u002Fli>\u003C\u002Ful>\u003Cp>If the agent can’t survive that, it probably shouldn’t be in your paper.\u003C\u002Fp>\u003Ch2>6) The survey is useful because it shows where the next fight will be\u003C\u002Fh2>\u003Cblockquote>“We will publish results from this experiment in the future.”\u003C\u002Fblockquote>\u003Cp>Anthropic says this survey is the baseline wave of a larger study, including a randomized experiment that gives researchers access to \u003Ca href=\"\u002Ftag\u002Fclaude-code\">Claude Code\u003C\u002Fa>. That’s the important part. The survey tells us who is already using these tools and what they think. The experiment will tell us whether the tools actually change productivity in a measurable way.\u003C\u002Fp>\u003Cp>What this actually means is that we’re still early. Right now, the field is arguing from anecdotes, self-selection, and partial adoption. The better question is not whether coding agents can help. They obviously can. The question is where they help, who gets them, and what they do to the shape of research over time.\u003C\u002Fp>\u003Cp>I like this framing because it avoids the lazy extremes. It’s not “AI will write all social science now.” It’s not “nothing is changing.” It’s a narrower, more useful claim: the research workflow is being reassembled around tools that can execute code, and the effects are uneven.\u003C\u002Fp>\u003Cp>How to apply it: if you’re building your own workflow, stop asking whether an agent should do everything. Ask where it can remove the most friction with the least epistemic risk. That’s the real design problem.\u003C\u002Fp>\u003Ch2>The template you can copy\u003C\u002Fh2>\u003Cpre>\u003Ccode># Coding-agent workflow for quantitative research\n\n## 1) Define the task\n- Research question:\n- Dataset:\n- Output expected:\n- Constraints:\n\n## 2) Give the agent the job\nUse this prompt:\n\n\"You are working inside a research repo. Your job is to:\n1. inspect the dataset and project files,\n2. write the analysis code,\n3. run the code,\n4. fix errors from actual runtime output,\n5. summarize every change you made.\n\nDo not invent variables, files, or results.\nIf something is unclear, ask before proceeding.\nReturn only the final code, the commands you ran, and a short explanation of the analysis choices.\"\n\n## 3) Force the workflow to stay honest\nBefore accepting output, require:\n- file paths used\n- commands run\n- errors encountered\n- fixes applied\n- assumptions made\n- final tables or figures generated\n\n## 4) Review checklist\n- Does the code run from a clean environment?\n- Are all variable names real?\n- Are the cleaning steps documented?\n- Are the results reproducible?\n- Would I be comfortable defending the analytic choices?\n\n## 5) Human-only steps\nKeep these out of the agent:\n- final claims\n- causal interpretation\n- paper framing\n- submission decisions\n- ethical review\n\n## 6) Good use cases\n- project scaffolding\n- data cleaning scripts\n- regression\u002Ftable generation\n- robustness checks\n- figure regeneration\n- appendix code cleanup\n\n## 7) Bad use cases\n- inventing methods\n- writing final conclusions without review\n- changing the research question midstream\n- hiding failed runs\n- replacing judgment with defaults\n\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>That’s the version I’d actually hand to a researcher or a lab manager. It keeps the agent in the parts of the workflow where speed helps, and it forces a human back into the parts where judgment still matters.\u003C\u002Fp>\u003Cp>Original source: \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fcoding-agents-social-sciences\">Anthropic, Coding agents in the social sciences\u003C\u002Fa>. My breakdown is original commentary built from that post, with the workflow template adapted from the patterns Anthropic describes and from practical research use.\u003C\u002Fp>","Anthropic’s survey shows where coding agents actually fit in social science work and gives a copyable workflow for using them well.","www.anthropic.com","https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fcoding-agents-social-sciences",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780315413286-4dxu.png","research","en","a655d074-7c36-424d-b63e-b0f9e31c708c",[17,18,19,20,21],"anthropic","coding agents","social science research","claude code","research workflow",[23,24,25],"Coding agents are different from chatbots because they can execute analysis, not just draft text.","Anthropic’s survey shows adoption is still limited and uneven across discipline, career stage, and gender.","The biggest near-term value is speeding up the front of the research pipeline, not replacing judgment at submission.",3,"2026-06-01T12:02:56.718646+00:00","2026-06-01T12:02:56.701+00:00","3103988e-c4fe-45e3-98ab-846500c9d507",{"tags":31,"relatedLang":43,"relatedPosts":47},[32,34,36,39,41],{"name":18,"slug":33},"coding-agents",{"name":19,"slug":35},"social-science-research",{"name":37,"slug":38},"Claude Code","claude-code",{"name":40,"slug":17},"Anthropic",{"name":21,"slug":42},"research-workflow",{"id":15,"slug":44,"title":45,"language":46},"anthropic-coding-agents-research-ops-zh","Anthropic 讓 coding agent 變研究 ops","zh",[48,54,60,66,72,78],{"id":49,"slug":50,"title":51,"cover_image":52,"image_url":52,"created_at":53,"category":13},"850449f2-e75b-4dbf-97c0-3590c6cbf097","crdts-keep-replicas-in-sync-without-locks-en","CRDTs keep replicas in sync without locks","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781011086602-cokl.png","2026-06-09T13:17:35.890527+00:00",{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":13},"7c6b6428-ba8d-4c59-840b-cf96a95139e5","post-deterministic-systems-autonomous-infra-en","Post-Deterministic Systems for Autonomous Infra","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781010190497-1grq.png","2026-06-09T13:02:33.235795+00:00",{"id":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":13},"53ec2203-e127-4bf8-8b3d-2dce8d156a54","causal-learnability-formal-language-tasks-en","Causal methods for measuring task learnability","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780987698514-ky8m.png","2026-06-09T06:47:35.103221+00:00",{"id":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":13},"55e7197e-f114-4b6c-b3e2-af1a3cd9dfa4","rl-training-hands-off-control-gradually-en","RL Training That Hands Off Control Gradually","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780986801034-gf8m.png","2026-06-09T06:32:33.516452+00:00",{"id":73,"slug":74,"title":75,"cover_image":76,"image_url":76,"created_at":77,"category":13},"93fc6735-b524-4baf-989f-645c4c47d593","omnigamearena-vlm-game-agent-benchmark-en","OmniGameArena benchmarks VLM game agents better","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780985895695-ugcj.png","2026-06-09T06:17:32.668876+00:00",{"id":79,"slug":80,"title":81,"cover_image":82,"image_url":82,"created_at":83,"category":13},"9f0c9505-6d75-411c-ba46-2382e8f295a5","turboquant-cuts-kv-cache-memory-6x-google-tests-en","TurboQuant cuts KV cache memory 6x in Google tests","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780906679116-fqdo.png","2026-06-08T08:17:22.276769+00:00",[85,90,95,100,105,110,115,120,125,130],{"id":86,"slug":87,"title":88,"created_at":89},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":131,"slug":132,"title":133,"created_at":134},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]