Coding agents have changed how we work as bioinformaticians. The change is recent enough that most published guidance is already out of date, and the tools are meaningfully different from six months ago.
The friction of writing and executing code is mostly gone. We now work most on things the AI agents can’t be trusted to do alone (yet). This includes:
- designing analyses
- constraining the agent’s decision-making capabilities
- checking the output
- interpreting results and
- recommending next steps.
We are able to do vastly more of these critical things because our time has been freed up from writing and debugging code. Below we’ve identified four tasks that keep us *very* busy.
1. Defining the question, then building the machinery to answer it.
We have found the best results when we specify the methods and ordering of analysis steps. Before the agent begins, we must set standing limits on what it gets to decide for itself. These practices prevent the agent from “going off the rails” with unnecessary or incorrect analysis.
2. Validating the data and keeping the provenance chain intact.
Data validation is a step both agent and human will often skip. No matter who (or what) is writing the code, it’s an essential step. Similarly, diligent tracking of the provenance of results, including recording exactly which dataset each analysis was run on, ensures the integrity and reproducibility of the analysis.
3. Interpreting and evaluating the biological meaning of results
Much of our result review time goes into evaluating whether a nice-looking p-value means anything biologically. We provide interpretation and evaluate whether the results are strong enough to support decision-making. We also assess whether additional analysis with complementary datasets is needed, and recommend next steps or additional experiments.
4. Determining if the client can stand behind the analysis.
Our job isn’t to make program decisions for our client. It’s to deliver work they can use to make those decisions. The reports must document assumptions, state limitations, and explain reasoning that someone who wasn’t in the room can follow. We ensure the analysis and the contents of the report support the decision-making needs of a particular project.
In a future post, we’ll talk about how an improperly-guided AI can produce inappropriate or incorrect results.
Coding agents are enhancing our work as bioinformatics consultants
As bioinformatics consultants, much of our career has been centered around coding. It’s a powerful shift to focus on the work that drew us into the field in the first place–the scientific judgment and interpretation.
We’ve also found that our productivity has increased tremendously, and our client reports have improved in size, thoroughness, and polish.
We can turn over anywhere from two to ten times as many results now, and each of those results is better. Our clients are better informed and are moving faster because we can support them better.
We would be glad to discuss any questions you may have about how Diamond Age can help you with AI and bioinformatics.