When you set out to build an AI or agentic system, think about augmentation, not replacement. Once you internalize that, the systems you build look very different and, to be honest, simpler and more practical.
Start any AI system with a “replace the job function” mindset, and you almost always end up with something more complex and less useful than if you had started with augmentation in mind.
Take incident auto-remediation as an example. If the goal is to auto-detect and auto-fix everything with AI, the design goes all over the place with edge cases, safety checks, rollback logic, and so on. In most cases, you will end up shipping nothing useful.
Now, look at the same system with the mindset of augmenting your existing workflow, and it looks completely different. The AI can be leveraged to surface root causes, suggest a fix, and hand the final call to an engineer. Suddenly, the scope is bounded, the failure modes are manageable, and the system actually ships.
By the way, this pattern holds beyond incident response and extends to anything and everything you can imagine.
So, whenever you are unsure how to scope an agent, ask yourself what it looks like to make the human faster instead of making the human unnecessary. That single reframe usually produces the better system.
Hope this helps.