Hello Bioinformatics lovers,

Tommy here. Today we will talk about agentic workflow vs automatic workflow.

People keep asking me the same thing: should I run my analysis through Nextflow or Snakemake, or hand it to a Claude Code agent? The question assumes one wins.

Both win, at different jobs.

A workflow manager runs a fixed pipeline. You write the steps in Snakemake, Nextflow, or WDL once, and every input travels the same path through the same graph. Same input in, same output out, run after run.

This is what you want in production. Cell Ranger on 10,000 samples. nf-core/rnaseq across a consortium.

A workflow manager scales, versions its containers, and restarts from the step that failed five days into a run.

It will never read a QC report and decide a batch is garbage, catch two samples that got swapped, or adjust when the data stops matching the spec you wrote.

An agent works the other way. Claude Code with a bioinformatics MCP reads an output, picks the next step, and calls the tool. Sometimes it gets the call right.

Sometimes it runs DESeq2 on TPM values with full confidence, and DESeq2 needs raw counts, so the answer comes out wrong and the agent never flinches.

An agent earns its place when you explore a new dataset, triage a strange QC report, or write a one-off comparison nobody will run again. It turns a vague question into a figure by Friday.

It struggles the moment you need the same thing done the same way twice, or anything audited, or anything where one wrong parameter changes the answer in silence.

Here is how I split the work. A 200-sample scRNA-seq project headed for a paper goes to nf-core/scrnaseq, same path for every sample.

A public dataset I want to understand by Friday goes to Claude Code in a scratch notebook I throw away after. Clinical variant calling goes to a Nextflow pipeline, locked and signed off.

An agent does not replace a workflow manager, and a workflow manager does not replace an agent.

If your analysis is solved, automate it. If your question is still moving, give an agent room to think. Most real projects need both, in that order.

What are you reaching for an agent to do that a pipeline would handle better, or the reverse? Hit reply and tell me where the line falls in your work.

Happy Learning!

Tommy aka crazyhottommy

PS:

If you want to learn Bioinformatics, there are four ways that I can help:

1. My free YouTube Chatomics channel, make sure you subscribe to it.
2. I have many resources collected on my github here.
3. I have been writing blog posts for over 10 years https://divingintogeneticsandgenomics.com/​
4. Lastly, I post daily on Linkedin and you may find gems there https://www.linkedin.com/in/%F0%9F%8E%AF-ming-tommy-tang-40650014/recent-activity/all/​

Stay awesome!