A field guide to AI + Science workbenches

This piece is in progress. It’s published here as a working preview so the structure and method are visible before the full draft lands — the analysis below is a scaffold, not a finished verdict.

A new class of software has appeared around the practice of research: the AI + Science workbench. These tools promise to sit between the scientist and the work — reading the literature, drafting analyses, running code, and keeping a record of how a result was reached. The category is young, the marketing is loud, and the boundaries between products are genuinely blurry.

This is a field guide, not a leaderboard. The goal is to map where each tool actually fits, what they share, and — most usefully — where the real gaps still are for a working lab.

What a “workbench” is trying to be

The pitch is consistent across the category: reduce the distance between a research question and a defensible answer, without sacrificing the traceability that makes the answer science rather than a guess. Three capabilities tend to define the class:

• Grounded synthesis — pulling from the literature and your own data with citations that resolve to a real source.
• Executable analysis — generating and running code, so a figure traces back to the exact steps that produced it.
• Provenance by default — a durable record of the reasoning, code, and conversation behind each result.

The contenders

We’re assessing the platforms hands-on rather than from documentation. The comparison will cover positioning, strengths, and gaps for tools including Claude Science and Benchling AI, alongside the lighter-weight synthesis tools researchers already reach for. Detailed, sourced write-ups per tool are the next section to land here.

The most interesting differences aren’t in what these tools can do — it’s in what they refuse to hide.

Where the real gaps are

The early read: the category is strong on drafting and synthesis, and still thin on the unglamorous parts of rigor — reproducibility across environments, provenance that survives a hand-off, and honest treatment of uncertainty. Those gaps are exactly where a partner who knows the science, not just the tool, earns their keep. That connection back to the three pillars is the point of writing this at all.

Check back as the per-tool sections and a summary comparison table are added.