Clojurians Log v2

Thanks. Yes, that was a very good summary!

I use it all the time. What about setup? I have some canned prompts that spawns off subagents like crazy. 😃

Do the subagents invocations count as github copilot pro requests, or is it one request only?

An example. Once I’va had some discussion with the agent so that it knows what I want to change I find that if it several agents are involved in the planning, I can get those more solid. So I do things like this:

1. Please gather information about how things are implemented, so that you can plan the changes. Use the repl, code, documentation, to verify assumptions.
2. Please create a new plan document for the implementation of this.
    Make it explicit in the plan that the implementation should _only_ implement and concern with what's in the plan, and that no other changes are allowed without first discussing them with the human developer (should something arise along the way that challenges the assumption that the plan can be implemented exactly as stated).
3. When you have created the plan, please use three parallel system-expert subagents to review the plan. Include the original ask from me, so that the reviewers can find misunderstandings that we may have missed. Tell the reviewers to use the repl, code, documentation, to verify assumptions.
4. Use three parallel system-expert subagents to cross-grade a different set of two reviews
5. Synthesize and summarize the synthesis to the chat
6. Update the plan with what you've learnt from the reviews and cross-grading

I’m don’t really know how premium requests are counted…

I probably should update awesome-copilot with some of my custom agents and stuff. But things are changing so fast, and I am still learning. Often still learning the old thing and not being able to keep up, so I don’t know how ready my prompts are for awesome-copilot.

What are you using to drive that? Last I looked (yesterday!), VS Code's Copilot can't directly spawn and orchestrate multiple agents, even if a specific model creates subagents as part of its work...

I just tell that to Copilot in the chat.

So the orchestrator in this case is the agent I am talking to in the chat.

I have also built a subagent dispatch thing in Joyride where I orchestrate using Clojure. But the same day I was sort of done with my MVP of that, the VS Code team released subagents, so I don’t use my Joyride system for this all that much. Still, sometimes that is useful, because some tasks are just better to prepare and orchestrate in code rather than have an indeterministic lunatic keep things together.

Ah, OK. So I'm already doing that I guess.

Full multi-agent would allow for different models for each agent.

You can do that by creating copies of the custom agents, using different models.

I guess Joyride could help with creating those copies on the fly.

I'll just wait until MS add native support in the Copilot Chat extension 🙂

I’ve tried to push them in that direction for some months now, but they don’t seem all that interested. We’ll see: https://x.com/pappapez/status/2021516302628569347

@lsolbach btw, Copilot has a tool for asking the human questions mid-session now, askQuestions. So I am going to retire the Human Intelligence tool some day soon. I may come back to it with a more complete human intelligence tool, but right now the overlap is pretty 100% (but not quite, the built-in tool does not have an auto-reply timeout.)

So here you are using them for redundancy and coverage. 🤔

I really have to get familiar with it.

I also use it for edits and running tests and stuff like that. Especially edits can get messy and the agent spends token window and focus on figuring out how to close a bracket or quote a string. When this happens in a subagent it doesn’t cause the whole task to fail. It just gets a bit slower, and perhaps more expensive. So I have in the basic copilot instructions, as well as in the custom agents that they never edit themselves. So they always delegate edits to a subagent.

Have to watch this: https://www.youtube.com/watch?v=GMAoTeD9siU