Loading request...
Users need the ability for agents to hot reload tools, where the agent writes the code for a tool from a skill file once and only rewrites it when absolutely necessary. This would balance flexibility and efficiency.
Keep SKILL.md body under 500 lines for optimal performance... Whenever I’m working with something new, I want to know HOW it works! Like npx do-the-thing - everything just works...well until it doesn't!! The problem with pre-built tool commands is they hide the good stuff that allows us to know what's really going on. When something breaks, you’re left staring at an error with no real idea what actually happened. So I've always been curious and watch the network calls, look at logs, dig into the source code if I can find it. Reverse engineer stuff... I’ve been noticing the same thing with agentic coding. What I've found is that it's far more powerful to create skills that contain scripts to do the work (PowerShell etc.) This way the agent can inspect the logic, see the assumptions, trace what ran, and even modify the script before retrying if something didn't work. Clearly there is a place for tool calling where it's so complex and a discrete complex units of work - but giving a skill a folder of scripts source code rather than pre-compiled tools has been giving good results. I'm still experimenting lots with this and trying to find the balance. What have you found? Here is the best practices advice when writing agent skills: https://lnkd.in/gx_iXPJa #GitHubCopilot #AgenticEngineering #VibeEngineering #VibeCoding #PowerPlatform