If My Agent Can't See Your Catalog, Will It Exist Much Longer?

Another edition of my dabbling with AI.  Somewhat like what I reported before - a customer asked if Ginkgo Cloud Lab's Echo-MS option could configure an assay to replicate the analytical results in a published paper, so I'm using Claude to do the heavy lifting.  Plus part of what I described earlier showed some issues on review.  But the big thing I'm keying into this morning is Claude is having difficulty with accessing a major chemical catalog I pointed it at - and I wonder how long that situation will be tenable.  Maybe not for a big company, but how many small companies will have an existential crisis when it turns out AI co-scientists and full agentic scientists have been silently passing them by?

The whole scientific world is scrambling to figure out how the machine learning (aka AI) revolution will positively and negatively impact the conduct of scientific research.  One large branch of this space are various tools to assist in designing experiments, ranging from chatbot sessions to identify the right equipment and consumables through emerging efforts at "agentic scientists" or "scientist agents" which try to evaluate existing data, form new scientific hypotheses, and the run bioinformatic or wetlab experiments. I wrote about one example of the such an agent-as-scientist, a collaboration between Ginkgo and OpenAI which optimized a Cell Free Protein Synthesis (CFPS) reaction. Ginkgo Cloud Labs is very interested in enabling such agentic scientists - please give your scientists a very large budget! 😀 We're definitely a work-in-progress - right now we've just marked up the webpages to try to make them maximally discoverable, but fancier approaches such as a full-blown Application Programming Interface (API) or Model Context Protocol (MCP) server are being explored.

These agents, whether just assisting or trying to act like full scientists, run into all the same issues we humans do - and then sometimes more.  

For example, I gave Claude the URL for the paper's HTML version and it immediately said it was paywalled.  Ain't for me - there's something being misinterpreted in the mess of templated HTML coming back from the website.  But it found the PDF so that worked.

In my case, as noted before, I'm trying to develop out a whole workflow - digest a paper or preprint, extract the compounds, compute (if needed) unique identifiers, find pricing information, and identify characteristic ion species to configure an Echo-MS assay around.

I mentioned in the prior item that in one case my search (then running on Gemini) claimed a compound could be purchased from Sigma-Aldrich, but if I looked at it the compound is discontinued.  Human saw a pop-up that AI wasn't initially tuned for, but then Gemini found it when I remarked on the discrepancy. 

Today the problem is a straight-up timeout, though apparently Claude then found other sites to get my quotes from.  Can I trust those as much?  And if my AI can't see the prices, could it set up an order for me?

An update on my prior Gemini work, turns out when I tried to get quotes on some compounds the startup synthesis vendor kicked back one of my CAS numbers as bogus.  I should have checked!  And once I did, turned out another compound was mismapped to a CAS number.  Of course, it would be nice if the academic preprint had CAS numbers and SMILES strings instead of just drawing of the compounds - I had Claude try to generate these but now I'm wondering how I would properly validate them?  Feed them into a drawing program & eyeball the structures for matches.  Definitely not my strong suit.  Perhaps have Claude and Gemini generate them & then check their own and the other AI's work?

But back to my current glitch.  If you are selling anything in the scientific space - contract research services, reagents, labware, equipment, furniture, consulting, data analysis services, et cetera - if your catalog isn't reliably accessible to the modern AI tools then you will lose business, and worse you will have absolutely zero visibility that these tools are driving right past your internet storefront.  Only if some loudmouth shouts it from the rooftops will you have any clue - and that's probably only if you have something very special the kvetcher knows about and so realizes their AI didn't find said item or service.  

I'm sure early in the e-commerce revolution it was the same thing - many sellers were stuck in their ways and had to be dragged kicking-and-screaming either to web storefronts or were dragged in a defeated mood to liquidation auctions.   There were certainly others that engaged in high profile hires, splashy PR - and dismal customer experiences.  In this case, I'd suggest that any large company should have someone dedicated to thinking about how to ensure their e-commerce site is always accessible.  Indubitably there will spring up consulting companies - and then plenty of spam in my email box - of small shops and consultants who will critique your offerings (this site gets a few a month about Search Engine Optimization, which are categorically ignored).  

In Lewis Carroll's Through the Looking Glass and What Alice Found There, the Red Queen tells Alice that they must run as fast as possible to simply stay in one place.  I'm sure e-commerce technology feels that way to sellers - but the Red Queen was on firm land and you are instead swimming in a virtual ocean with powerful tides.  It's sink or swim - and clearly any commercial venture would rather swim.