-
Notifications
You must be signed in to change notification settings - Fork 1.1k
Description
Describe the feature or problem you'd like to solve
Make it easy to provide only the applicable context to agents, especially if you want to support many tools and skills but only as needed
Proposed solution
IIRC, every time I launch copilot all MCPs and plugins are always on, so that's a bunch of context used up even if some MCPs are very task-specific. In the case of my team, they are trying to push for using copilot cli for everything work related with tools like workiq and Azure DevOps (we're still transitioning to GH, don't judge) to help manage workflow and keep agents in sync with chats. The problem is, once I've picked up the next task, I don't need these MCPs until this task is finished. It's just clogging up my context. And if I'm just running an agent/skill that just needs to do a simple coding task (e.g., make sure the code matches team preferences according to this doc) then having access to all these tools is overkill and is going to dilute the model's focus.
Much better would be to have a way to configure that this is a "work" agent that has access to all of my tasks etc. and those MCPs vs a "coding" agent which just has access to MCPs/tools that help with getting coding done.
Example prompts or workflows
- Typical usage: I need to work on coding project. I want only coding related MCPs and skill context filling up my agent's precious context
- I want to get a summary of recent conversations to distill action items: have an agent that can use
workiqand ADO I don't need it to worry about how to use Playwright and GH - When planning out features or documenting the codebase, I need as much context as possible for the code at hand not how to use irrelevant skills/MCPs
How could we configure it?
- Manual
- User can define some custom agents with limited tool/skill use or explicit tools/skills to ignore and if the tool/skill doesn't make the cut it doesn't go into context. I believe this already exists to some degree? Maybe the problem is that there is no way to do this for the generic agent?
- User could mark MCPs with some sort of tag (e.g., task vs coding vs review) and then also mark agents with tags and if both are aligned then we apply the tool to the agent.
- Automatic - we force mcp/skills to have a limited preamble (max 256 characters or something) that explains when the agent would benefit from using it, and the agent can expand to see the whole skill as needed, but we don't by default dump everything you need to know about every possible skill by default.
Additional context
No response