Rule #4 is where I see the handoff break. The skills get written during build, capture domain knowledge and edge cases - then six months later a different team inherits the agent and has to modify the skills without understanding why they were structured that way.
'Avoid infrequent events like signed in' makes sense to the original builder. To the inheritor, it's just a constraint they don't know whether to keep or remove.
this resonates with my learnings building agent(s) too! agent should be able to see what user sees, we need to provide opinionated defaults, and a few other things. (https://harshalpatil.substack.com/p/ai-copilot-expectations-day-one)
I found it interesting that you moved from UI to CLI, whereas my efforts were to move from CLI to UI so that our observability matches the user view.
Most systems do not fail because the agent is wrong. They fail because the system does not know what to do when the agent is partially right.
How are you handling incomplete workflows?
Rule #4 is where I see the handoff break. The skills get written during build, capture domain knowledge and edge cases - then six months later a different team inherits the agent and has to modify the skills without understanding why they were structured that way.
'Avoid infrequent events like signed in' makes sense to the original builder. To the inheritor, it's just a constraint they don't know whether to keep or remove.