A few ways Claude actually fits into research

In a companion post I argued that vibe research, the mode where you direct and an AI agent executes, is worth being optimistic about. That one stayed up at the level of ideas. This one comes back down to the ground: what does an AI agent like Claude actually do inside a research workflow, concretely, today?

Before the examples, one boundary that's worth holding onto. Agents are genuinely strong at some things and genuinely weak at others, and the line between them isn't where you'd guess. They're fast, they don't get bored, and they'll follow a rigorous procedure step by step without cutting corners. What they lack is original judgement and tacit field knowledge, the sense of which question is alive and which is exhausted. Zhang (2026) puts it as four words I haven't been able to shake: coverage is not comprehension. Keep that in mind as you read what follows. Everything below plays to the strengths and leaves the judgement with you.

1. Literature work you can actually audit

Start with the most obvious job: reading and synthesising the literature. This is where most researchers already reach for AI, and it's also where the failure modes are easiest to demonstrate. Fabricated citations. Summaries of papers that don't quite say what the summary claims. The usual.

What changes the picture isn't a smarter model so much as a workflow that leaves a trail. The interesting design here is retrieval-augmented research, where the agent doesn't generate claims from memory but pulls actual sources, embeds and searches them, and ties every statement back to where it came from. Later I might make a post about creating an LLM wiki for research purposes.

A tool called Notez Nerd is the cleanest example of the shape I have in mind. It's local-first, and it's built around traceability. As it works through query analysis, literature collection, extraction, and synthesis, it keeps research logs, citation traces, and a record of every tool call (Notez Nerd, 2026). The agent will still make mistakes. What matters is that you can find them, because the path from claim to source is right there. The agent does the ploughing through papers. You keep an audit trail you can actually check later.

2. A pipeline with the brakes built in

The example I find most convincing, partly because it's literally built on Claude, is a Claude Code plugin called Scholar-Skill.

Here's why it's interesting. Rather than handing the whole study to one agent and hoping, it breaks a social-science research pipeline into 26 specialist skills spread across 18 phases, and it wires in 53 quality gates along the way. Most of those gates the agent can clear on its own. Five of them are hard stops, where the workflow refuses to continue without a human: data safety, literature and theory verification, a pre-draft review, citation verification, and ethics compliance (Zhang, 2026).

That structure is the opposite of the "accept all" reflex. It's an agent on a leash, by design. Zhang (2026) is explicit that it isn't an autonomous research engine. It produces options, and a researcher still has to evaluate them, choose, and revise. The orchestration is quite impressive and keeps the human in the loop throughout the process.

This is the template I'd want for serious work. Let the agent run fast across the codifiable parts. Force a stop at the points where a quiet error would do real damage. Verification stops being something you remember to do and becomes something the workflow won't skip. Of course you could do something similar with skills like /goal and /loop, where you describe the multistep workflow and clearly state the stopping points where human consent is required.

3. The reality check: agents aren't good at checking yet

The task you'd most love to hand an agent is verification itself. Take a paper and its code, rerun it, and tell me whether the results actually hold up. REPRO-BENCH tried exactly this, on 112 real social-science papers, each with a reproduction package and a set of headline findings. The agent gets the PDF, the code and data, and the claims, and has to score how well the work reproduces (Hu et al., 2025).

The agents struggled. The best off-the-shelf system scored around 21 percent, below random guessing. A purpose-built workflow pushed that to roughly 37 percent (Hu et al., 2025). Still nowhere near reliable. They trip on unglamorous things: finding the right script in a messy package, parsing Stata logs, installing dependencies, resolving file paths, noticing a subtle mismatch between what was reported and what came out.

Things may have moved on since (I'm writing this in June 2026). One habit that works well for me in coding is keeping review in a clean session: I make my commits and PRs, then open a fresh session with Claude and ask it to review and simplify the work, usually with multiple agents to improve quality. The same split would suit a research paper: draft in one session, review and verify in another.

4. Domain copilots, with appropriate salt

In biology, Owkin's K Navigator is pitched as a copilot that lets scientists "talk to their data as easily as they talk to a colleague," handling literature scanning, idea generation, and early analysis, with productivity claims in the range of twenty times (Owkin, 2025).

I don't take the multiplier literally. Vendor numbers are vendor numbers. What I do take seriously is the form: a domain-specific agent that knows the shape of one field's data and tooling, sitting next to a scientist rather than in place of them. Generic agents are useful everywhere. Agents that understand genomics, or materials, or whatever you work in, are going to be useful in a deeper way. That's the version of this I'm most curious to watch.

How I'd actually use it

Underneath the examples, the working pattern is plain. Let Claude execute the parts that are labour rather than insight. Keep yourself in the loop at the decisions that matter. Make the whole thing auditable, so you can always trace what the agent did and check it.

The agent gives you back time and reach. The judgement and the accountability stay yours.

Sources

Hu, C., Zhang, L., Lim, Y., Wadhwani, A., Peters, A., & Kang, D. (2025). REPRO-Bench: Can agentic AI systems assess the reproducibility of social science research? arXiv. https://arxiv.org/abs/2507.18901

Notez Nerd. (2026, February 6). Vibe research 2026: The new paradigm of human-AI collaborative research. https://www.noteznerd.com/blog/vibe-research-2026

Owkin. (2025, June 16). Vibing with your research: The future of vibe research for biology. https://www.owkin.com/blogs-case-studies/vibing-with-your-research-the-future-of-vibe-research-for-biology

Zhang, Y. (2026). Vibe researching as wolf coming: Can AI agents with skills replace or augment social scientists? arXiv. https://arxiv.org/abs/2602.22401