Do your task agents deployed automatically when a new task appears? I really like your workflow, and would upskill mine.
Not usually, but sometimes.
Usually the backlog contains all sorts of stuff at various levels of thinking, or blocked by other things. Sometimes I would start a task agent to just "research and expand on this issue and update the GH with a comment"
The main flow in VK then is "read GH #123, research repo, and present plan as TDD". I may juggle 3-4 of those at a time working on different areas.
Other times, I would set up an agent in VK to run an epic "break down this issue into sub issues and add them all to GH" then, "sequentially use VK MCP sub tasks to assign subagents and start attempts one by one; review each as they reach review stage; merge into your worktree when completed."
A bit of manual handholding there that is scriptable when I get a round to it, but good for breaking down and delivery of bigger stuff.
E.g. started an epic as "make Clips a first class content category in Shosho", which expanded into a whole array of issues (subscriptions, context provider, stores, screens, playback, etc) then it just ran them all one by one until done, and now we have Clips.
I think the benefit of keeping the thinking in GH is if you ever need to eject from an agent session/attempt and pick up with a different model (or human) it's all documented in the issue, in a much tighter and easier to use way than scrolling back through the agent chat.
Still learning but this is working for me at the mo.
Probably will all be different by next month!
Usually the backlog contains all sorts of stuff at various levels of thinking, or blocked by other things. Sometimes I would start a task agent to just "research and expand on this issue and update the GH with a comment"
The main flow in VK then is "read GH #123, research repo, and present plan as TDD". I may juggle 3-4 of those at a time working on different areas.
Other times, I would set up an agent in VK to run an epic "break down this issue into sub issues and add them all to GH" then, "sequentially use VK MCP sub tasks to assign subagents and start attempts one by one; review each as they reach review stage; merge into your worktree when completed."
A bit of manual handholding there that is scriptable when I get a round to it, but good for breaking down and delivery of bigger stuff.
E.g. started an epic as "make Clips a first class content category in Shosho", which expanded into a whole array of issues (subscriptions, context provider, stores, screens, playback, etc) then it just ran them all one by one until done, and now we have Clips.
I think the benefit of keeping the thinking in GH is if you ever need to eject from an agent session/attempt and pick up with a different model (or human) it's all documented in the issue, in a much tighter and easier to use way than scrolling back through the agent chat.
Still learning but this is working for me at the mo.
Probably will all be different by next month!
1
