What I’d Ask AI Agents to Do, and What I’d Never Delegate
My everyday realistic guide to delegation in the world of AI
So, here’s a confession: I love automation.
Give me a well-trained agent that can save me 3 hours of grunt work, and I’ll light a candle in its name.
But lately, I’ve been thinking deeply about judgment, especially as AI agents get more powerful. It’s one thing to know that they can write your test files, clean up your code, and draft documentation. It’s another thing to know what not to let go of.
Because here’s what I’m learning:
Not every task is meant to be automated.
And not every decision should be outsourced.
What I am gladly letting Agents handle
Boilerplate Code & Repetitive Patterns - Writing DTOs, mapping interfaces, generating CRUD scaffolds - yes, please (don’t judge me). These tasks don’t need my creativity. They need accuracy and speed.
Test Stubs & Refactoring Suggestions - Let the agent scan my codebase and surface what needs cleanup or coverage. I’ll still review it, but I don’t need to spend time thinking about where the camelCase slipped into snake_case.
Summarising Long Threads & Docs - I don’t need to manually crawl through a 200-message Slack channel to know what the debate was about. Let the agent summarise it. I’ll verify what matters.
Converting JSON Payloads to Typed Structures - A task I’ve done more times than I can count, and never once enjoyed. Off you go, agent.
What I’d probably never delegate
Defining what to build - Product sense is sacred. It’s one thing to ask, “Can this be built?” It’s another to ask, “Should this even exist?”
AI can surface trends but it takes human context to feel real pain points.
Shaping Systems and Boundaries - Understanding what belongs together and what doesn’t, that’s not just about structure. It’s about clarity. About anticipating change. That’s still a job for thoughtful engineers.
Naming - Honestly, naming isn’t just semantics. Good naming forces clarity of thought. Have you ever lost control of your codebase? I mean, a codebase you wrote yourself but months later could not understand how parts of it work? Bad naming leads to long-term confusion. Agents don’t yet know why UserActivityTracker is better than EventLoggerX3.
Giving Feedback to People: I’ll delegate summarizing a PR. But when it comes to offering kind, constructive feedback? That requires emotional context, empathy, and presence. It’s a privilege, not a task so this is something I don’t delegate.
I strongly believe that the real skill today is knowing the difference.
The temptation now is to ask: “What can I automate?” But the better question is:
“What should I automate, and what should I stay close to?”
That’s what separates reactive engineers from strategic ones.
That’s what turns tooling into leverage, not just noise.
In this AI age, judgment is the new 10x.
And your greatest strength might just be knowing when to stay human.
What about you?
I’m curious, what’s one thing you’re now letting an AI tool or agent do for you?
And what’s one thing you’re holding onto very deliberately?
Let’s talk.