Agents amplify whatever model you already have

AI did not make engineering leadership easier. It made the cost of weak leadership visible faster.

This year I kept hearing the same question from engineers on AI teams:

Why do leaders trust the companies selling AI coding tools more than the judgment of their own teams?

AI really does speed up development. The undersold cost: code changes rise faster than the team’s ability to stay oriented. Without discipline and human process, people do not feel “10x.” They feel overwhelmed and quietly frustrated.

What kept my team sane:

→ My job is context, constraints, and judgment — not typing faster than Cursor → 1:1s hold trust and hard news. Agents cannot replace the conversation → ADRs/RFCs are the shared brain. Skills encode how, not why → Architecture only helps if the team co-owns it → Shared skills compound learning. Private prompt stashes do not → Automate mechanical review. Humans protect intent — and pause PRs that smuggle architecture

The operating model:

judgment → architecture → enforcement → automation 1:1 → ADR / RFC → code review → skill

Agents amplify whatever model you already have. Clear fundamentals make you faster. Fuzzy fundamentals make the fuzziness ship.

Full write-up: https://ai.ksopyla.com/posts/what-i-learned-leading-ai-engineering-team/

If you lead an AI team: where does your engineering discipline live now that agents read the repo too?

#AIEngineering #AILeadership #SoftwareDesign