Would you let AI agents deploy code at 3 a.m. without you? That question reveals where humans belong in the loop. Here’s my three-part system for deciding what to delegate.
The story
If your AI agents shipped code to prod at 3 a.m. while you slept, would you be cool with that?
That question is how I decide *where* a human belongs in the loop.
In my agentic coding teams, I split it three ways:
1. Before-the-loop: I design the game board
I stay “around the loop” defining goals, constraints, and boundaries.
Things I decide up front:
– Which tools agents can call
– Which systems or data they can touch
– Where they are never allowed to make irreversible changes
2. Human-in-the-loop: I approve the sharp objects
For high‑impact moves, agents propose; I approve.
Checkpoints like:
– Merging PRs or touching anything production‑ish
– Granting new repo, credential, or data access
– Architecture changes that affect security, cost, or compliance
3. AI-in-the-loop: I lead, they assist
For code review, architecture, prioritization, I stay the primary decision‑maker.
Agents surface options, risks, and edge cases; I choose.
To avoid automation bias, I treat agent output as proposals, not truths, and reserve full autonomy for low‑risk, reversible work: refactors, boilerplate, tests, scaffolding.
The goal is augmented automation, not blind delegation:
More mechanical work on the agents, more intent and judgment on me.
My gut-check for every action:
“Would I be comfortable if this happened without me ever seeing it?”
If not, that’s where I put a human squarely in the loop.
Recent Comments