We’ve started to play with the idea of Claude Code running Nextflow modules in dynamic workflows (as opposite to static DAG-based pipelines). This might be especially relevant for microscopy image analysis, if we can define well-specified objectives. Maybe even loosely defined objectives evaluated by vision-language models (VLMs) could work. The idea would be to automate getting to a reasonable processed output, then a human can take over and tune each step further if necessary. One issue with using nf-core modules for this is that module parameters are typically not specified at the module level. Rather there are pipeline parameters (

nextflow_schema.json

) which are mapped/injected into modules in pipeline-level configs. Some modules specify a documentation link (e.g. https://github.com/nf-core/modules/blob/master/modules/nf-core/admixture/meta.yml) that the agent could read. But this link can be anything and it’s not clear how easy it would be to have an agent ingest this cleanly. Another idea might be to add a

README.md

(or

AGENT.md

?) into each module that describes not only the CLI parameters, but how to run the module and what to expect from it. I was curious if anyone else is looking into this, or has any thoughts along these lines. @Tatiana Woller @Tong LI @Cristian Goina @Phil Ewels

Yes, we are 😊

Yeah, I remember that demo! What I’m talking about is a bit different, for the agent to orchestrate modules from the outside.

Do you mean running the analysis using an AI or letting the AI create a workflow for you here?

The former. You can already use AI to create workflows, for sure. But the end point of that is still a static pipeline which can't easily adapt to the data being processed. If the agent could instead just run the modules directly, then it could iterate any given step with different parameters until it has a "good" looking result (defining that might be the hardest part).

Interesting approach! How will you ensure the reproducibility of this setup?

Most of the traces should be logged in

nextlofw log

etc and we can also ask the agent to write the configs used for all runs, right? I see it more like a

hyper-parameter search

step, which is often the most time-consuming for image analysis. @Konrad Rokicki may have a different use case though