As suggested above, I've rarely been able to sustain such a coordinated effort on manuals - I've always been a "let's dive in and hope nothing breaks permanently" sort of guy. But now we are in a world of AI agents, and I'm so glad since one of my regular pursuits now involves technology exponentially more complex than that PC and also lacking a manual: Ginkgo's autonomous laboratory installation Nebula.
For both reasons of professional pride and because promoting the tech is my job, I am trying to build a working laboratory protocol on the system. For that we have agents to assist. Today I'm focusing on an aspect of the agent for which my appreciation continues to grow: the agent as master guide to Nebula.
Nebula consists of over 100 of our RAC (Reconfigurable Automation Cart) units. Most RACs have a single scientific instrument; some have gangs of 2, 4 or 6. Something like 120 instruments of 30ish different types on the system. Making those numbers precise is not particularly useful; the Reconfigurable and Automation components of the acronym mean the system can change radically in very short order. I had the experience one day of giving a tour intro spiel in one spot and finding that occupied by new RACs before the group exited the lab. Haven't caught one mitosing yet, but I wouldn't rule it out.
So that means any manual - even if in Wiki form - will be hard to keep up. There's also what is currently active versus what might be down for maintenance. There's also the matter that some instrument types have variations. For example, only some of the Labcyte Echo instruments are kitted out to handle low dead volume plates.
Our goal is for our Catalyst software to support all the operations an instrument can execute, but there could be discrepancies. That's a hard task and it is better to mostly support an instrument than not have it at all. Instrument makers may not surface all operations through their API.
So even if I have a manual - say pulled from the Internet - it might not reflect our specific instance or configuration, or there might be "check your model number, software revision, etc" or other ambiguity.
But we have an agent! So I can ask Catalyst Agent these questions. Plus I can have it teach me the difference between multiple instrument types that do similar things. For example, we have two different plate reader types on Nebula to support optical measurements. Three different bulk reagent dispenser types. Three different instruments for moving liquids between plates. How are they technically different in terms of what assays they can support? How are they operationally different? That my agent can teach me.
Critically, the agent can tutor me specifically in the context of whatever protocol I am trying to build and provide me with the specific tradeoffs that are relevant to what I am doing. What are the timing differences? What small features - such as onboard shaking - are available? Which device is better for exactly what I want to do?
And if the agent does come upon gaps between the instrument's inherent capability and what our software exposes, it can automate my feature request submission! And volunteers to create and submit it! Our talented agents team must find that a double-edged sword - feedback submitted in a consistent way, but they've enabled me to be an extremely squeaky wheel.
There's much debate about how to use AI in intellectual pursuits; where does it speed discovery and where does its use dull the mind. In my opinion, having the agent know Nebula inside-and-out, up-to-the-minute, is absolutely a win. I don't absolutely need it for giving a general introduction to Nebula components on our tours (sign up for one!). But scheming up experiments to run on Nebula? For that I lean hard on my agentic co-conspirator.
No comments:
Post a Comment