Laurie Voss: What's Left for Engineers When AI Writes the Code?
Laurie Voss, co-founder of npm and head of DevRel at Arize AI, gave a talk titled “AI Writes the Code. So What's Left for Engineers?” It overlaps with a lot of what I've talked about this year. The interesting part is where we draw different conclusions about the future of engineering from the same diagnosis.
Voss's reasoning is that agents have made writing code dramatically cheaper, and reviews can't keep up. Once we automate review, the bottleneck will move to operations and security. He thinks we'll master those too in five to ten years. People will be left to figure out what others need, define a good outcome, and turn it into a product. He calls that person a product engineer.
We agree on the diagnosis.
In my HighLoad++ talk, I looked at the same problem. Speeding up individual roles leaves the queues and waste between them in place. You can end up with more requirements, code, and tests while delivering about the same number of useful changes to the user. That's why you have to look at the task's entire journey, including waiting, rework, and release.
I also like his practical advice to get experienced engineers' knowledge out of their heads and into specifications, rules, and evals. If the same comment keeps coming up in review, can you make it a check the agent receives while it's still working?
The story with junior engineers is familiar too. Evgeny Sergeev, Alexey Litvinov, and I discussed entering the profession without easy tasks. I suggested assessing how a newcomer framed the task, evaluated the plan, and explained the result. A working feature now tells you far less about what its author learned.
Here's where I'd push back.
Voss sees understanding the need as the last human bottleneck. The details of a particular bakery are in the owner's head and aren't in the model's weights. That doesn't mean an agent can't find them out. It can ask questions, observe the process, and test hypotheses too. Voss candidly notes that he may simply be treating the part he enjoys as irreplaceable.
I looked at this objection in my September long read, “When Agents Write the Code: What's Still Engineering?” We can't build a profession on the hope that machines will never learn to understand a task. I care more about who sets the goal, which actions are permitted, and how we will check the consequences of a decision.
That's where our views on autonomy diverge. Voss ends by suggesting that we focus manual review on code with a large blast radius and gradually hand the rest to automated reviewers. I'd add some conditions: independent verification, a clear scope for allowed changes, failure detection, and tested recovery. A small diff can cause a big problem. Rolling back a version can't unsend a message or take back data that has already been disclosed.
That fork in the argument is worth watching the talk for. Are we preparing engineers to find the next task AI hasn't mastered, or to build a process in which we have good reasons to trust more capable agents with more work?
#AI4SDLC #AI #Agents #Engineering #Management #Learning