I’ve been experimenting with a way of running repeatable tasks with AI that goes a bit like this:
- Discover the best way to execute the task by asking an agent to install whatever tools it needs (MCPs, local scripts, SKILLs, etc.).
- Ask the agent to execute the task. Review the output and repeat he request asking for improvements wherever needed.
- Ask the agent to create a new SKILL that reflects all the steps it had just done.
At this point I can simply run the SKILL whenever I want to execute the same task again. I use for some time, asking the agent to improve it until I’m fully satisfied.
- Ask the agent to convert the SKILL into an Arazzo document.
This is the step that I’m still experimenting with and I’m not fully satisfied with the result.