Discussion

Last updated on 2026-07-14 | Edit this page

Where this goes next

This lesson is a starting point. A few directions worth discussing as a group or exploring on your own:

  • Central ePIC infrastructure. Instead of every collaborator running their own MCP servers, the collaboration could host shared servers and MCP proxies, a vetted library of skills, and agent harnesses that anyone can reuse. What belongs centrally, and what stays local?
  • More tools, same pattern. This lesson already ran three servers — uproot for analysis, rucio and xrootd for data discovery — and Episode 6 catalogs documentation search. What other ePIC workflows — calibration lookups, run planning, detector conditions — deserve a server next?
  • Trust and reproducibility at scale. The trust checklist (check the result, make it repeatable, keep a record, watch the cost, stay in the loop) matters more, not less, as analyses get more automated. How do we make an automatic result as solid as a hand-made one?
  • Cost and openness. The assistant table on the Setup page will keep changing. Open tools (opencode, Cline) plus local models, or commercial free tiers — what is the right default for new collaborators?

Further reading