Human-in-the-loop agents

We talk a lot about AI taking over tasks. The reality is messier. She works best when we don’t let her run wild, but also don’t keep her on a leash so short she trips. The sweet spot is us guiding, her acting.

Hybrid workflows

Think of it like pair programming. She writes most of the code, we stop her when she’s headed for the cliff. That means she does the routine bits fast while we keep the judgment calls. The result is a split workflow: she handles the parts that bore us, we handle the parts that break things if wrong.

Human approval

The magic isn’t in speed alone. It’s in deciding when she needs a hand-raise. Maybe she drafts an email, but it doesn’t go out until we hit “send.” Or she books travel, but the final itinerary waits for our nod. These pauses feel small, but they’re the guardrails that keep her useful.

Guiding autonomy

We can’t pretend she’ll know when to stop. That’s our job. A simple rule of thumb: let her roam where mistakes cost little, rein her in when mistakes hurt. Think: sorting files = safe. Approving invoices = not safe. It’s not fancy, but it works.

The loop itself

Here’s the loop: she acts, we check, she learns. Then the next time, she asks less often. Over time the loop shrinks because she gets better. But the loop never vanishes. If it does, we stop being partners and start being bystanders.

We’re coders. We like autonomy too. Watching her stumble reminds us: even agents need a teammate. And that’s us.