Let Your Agents Review, But Look Before Humans Do
Agents can run multiple rounds of review and verification before you look. But ship to humans without looking yourself, and something subtle always feels off — because fully specifying a system that feels right to a human is hard.
Some interesting thoughts in Noah Smith's piece, "Your future job will be to keep AI on task."
I've had a lot of luck letting different agents do multiple rounds of review and verification before I look at things — but I've regretted it when I don't look at things at all before other humans use the system I'm working on, even with clear acceptance criteria checked off explicitly.
There's always something in the details that is a little off, mostly because it's hard to completely specify a system that feels right to a human, at least while humans are the target user. There's lots of software where agents are the target user and that's a different story. Even then, the agents are still working on behalf of something humans want.
Related insights
Vibe Code to Explore, Engineer to Expand
Vibe coding is perfect when you're still finding out if anyone wants your thing. Once real usage and real money show up, agents make software engineering matter more, not less.
Sharing Compiled Context Across a Team Is Still Too Hard
An agent-first, human-friendly knowledge base is one of the most impactful tools in my day to day work, yet sharing and maintaining that context across a team is still surprisingly hard.
The Next Phase of AI Security Is Going to Be Less Fun
More people get to build software than ever, but the attention now required for security and breach recovery is draining a lot of the joy, especially for teams with legacy systems to bring up to date.
Get More Like This
Follow along as I build and share what I learn
Found this helpful? Share it with your network!