Planned support for global/user-level custom agents in Positron Assistant

[connormfrench]

I love the new custom agents feature, but the need to copy my agents installed to ~/.claude/agents/ or ~/.copilot/agents/ is a bit cumbersome, especially for agents I use across many projects. Are there plans for detecting user-level/global agents files for access in Positron Assitant?

[jthomasmock]

Thanks @connormfrench !

Can you share a bit more about how you are using /agents/in general and are you intending to use them with Positron Assistant directly?

Are you looking for Positron to automatically discover subagents and other customizations in other directories?

Also are you planning to continue focusing your use of agents or move to more skill.md based approach relative to Positron and it's Assistant?

Thanks, and happy to discuss further.

[connormfrench]

Hi @jthomasmock!

Currently I use custom agents to work more effectively with development/documentation tools and navigate my user-level environment in a directed manner. For instance, I created a custom revealjs agent to setup and modify new Quarto revealjs presentations within my organization, which involves the agent copying a template and other custom materials (e.g. scss file, logo photos, etc.) from a directory on my machine into the project directory and follow guidance to setup or modify the presentation. I use this agent regularly across repositories and keep the .agent.md file in ~/.copilot/agents/ so I can access it in whatever project I'm working on. In VSCode, the Copilot Chat is able to find anything in ~/.copilot/agents/ or ~/.claude/agents/ (this is convenient for claude-installed agents) so I don't have to copy it into each project I'm working on.

You're correct- I'd like Positron to automatically discover subagents in other directories, specifically those associated with the LLM provider I'm using (~/.copilot or ~/.claude).

I find that most of my custom agent creation isn't project-specific, but more general-use across projects. Most of my project-specific direction goes in a .instructions.md file. I prefer using agents because I find that skill discovery by LLMs is still inconsistent, even with skills that tell the LLM to use skills. However, I'm finding a hybrid approach is pretty successful- I direct my agent to access specific skills when it needs to perform relevant tasks.

If it helps, I operate in the bioinformatics world, where projects are generally production-level (i.e. Nextflow/Dagster) pipelines, Shiny dashboards, python/R CLIs, or adhoc data analyses.

Thanks, I appreciate your reply!