Treating AI like a 10x engineer gets you confident garbage. Treating it like a supervised junior gets you leverage. Here’s the protocol that’s working: tight specs, role separation, brutal feedback loops, and humans owning architecture while agents handle implementation.
The story
What I just learned about building AI + human dev teams:
If I treat an AI assistant like a 10x senior engineer, I get clown code with great confidence.
If I treat it like a sharp junior on a tight leash, I get leverage.
The pattern that’s working for me:
– I own the architecture and intent. Domain model, boundaries, failure modes, that’s on me. The agent just explores implementations *inside* that box.
– I don’t ask, “Design the system.” I ask, “Fill in this function,” “Refactor this file for clarity,” “Draft tests for this behavior,” always in the context of the existing codebase.
– I use multiple roles: a coding agent, a testing agent, a refactor/review agent, and sometimes a docs agent. Same codebase, different jobs.
– I keep a tight loop:
1] I write the spec.
2] The agent proposes code.
3] I critique and iterate.
4] I run tests, inspect diffs, then decide what lands.
– When the AI is guessing, I make it say so: assumptions lists, citations, things I can verify before anything ships.
The real unlock: treating documentation, tests, and design notes as the *shared language* between humans and agents.
I’m not tracking “AI usage.” I’m tracking cycle time, defect rates, and how much more time humans spend on architecture instead of boilerplate.
Hybrid dev teams aren’t about replacing developers. They’re about moving us up the stack.
Recent Comments