Why human-in-the-loop
beats full autopilot.
Watch a fully autonomous coding agent for five minutes and you will want one. It reads the repo, forms a plan, edits a dozen files, runs the tests, and reports back like a tireless junior who never asks for a coffee break. It is genuinely impressive. It is also, most of the time, a demo.
The trouble with autopilot is not that the model is dumb. Models are good and getting better every quarter. The trouble is that "good on average" is a terrible property to hand the keys to. An agent that is right ninety-something percent of the time still means that every so often it deletes the wrong file, ships the wrong fix, or quietly wanders off-plan while you are looking at something else. And the failures are not evenly distributed. They cluster exactly where the work is ambiguous, which is exactly where you would have wanted a human to weigh in.
Autonomy is not the goal. Leverage is.
Somewhere along the way the industry decided that removing the human was the whole point. Fewer taps, less oversight, more magic. But nobody actually wants an unsupervised agent rewriting half of a project they care about. What people want is leverage: the AI does the typing, the reading, the boilerplate, the tedious plumbing, and they stay the one who decides what is worth doing.
Those are different things. You can have enormous leverage and still keep a firm hand on the wheel. The mistake is treating "the agent did it while I was away" as the win condition, when the real win is "the project moved forward and nothing happened that I did not want to happen."
The agent proposes, you dispose. The worst case becomes a tap to reject, not a production incident.
The unit of trust is a gate
Here is the bet we made when we built ShipItFam. The thing worth engineering is not a smarter, more autonomous agent. It is the gate: the moment where the agent stops, shows you what it wants to do, and waits for a yes.
Get the gate right and everything downstream gets calmer. The agent can be aggressive and creative because a bad idea costs you a single tap. You can hand it fuzzier goals because you will see the plan before it runs. And crucially, you decide how tight the leash is. Feeling cautious on a fragile part of the codebase? Approve every command it wants to run. Confident on a greenfield feature? Let it work and review the whole thing at the end. Same agent, same model. The only thing that changes is where the checkpoints sit, and you set that per project.
That is the difference between a co-pilot and a loose cannon. A co-pilot flies the plane and lets you take the controls at any moment. A loose cannon just goes off.
Asynchronous, not absent
The obvious objection: doesn't approving things slow you down? If oversight means sitting at a desk babysitting a terminal all day, then yes, it would defeat the purpose entirely.
So we moved the gate to where you already are. Each step the crew wants to take becomes a card on a board on your phone. You approve from the sofa, the supermarket queue, the back of a taxi. The agent keeps working through everything it can, and only the decisions that genuinely need you rise to the top. Oversight stops being a chair you are chained to and becomes a notification you glance at. Present enough to steer, absent enough to have a life.
What we will not automate
There is a short list of things a human should always own, and we treat it as a hard line rather than a setting. Nothing merges to your main branch on its own. Nothing deploys itself. The agent works on its own branches inside your workspace and pushes back only when you say so. Your main stays yours. If that sounds conservative for an AI product, good. The conservative part is the product.
None of this is anti-AI. We are enormous believers in what these agents can do, which is precisely why we want them running hard. A leash is not a muzzle. It is the thing that lets you let go of the collar, because you know you can pull it back in an instant.
Full autopilot optimizes for a demo. Human-in-the-loop optimizes for the thing you actually want: real work getting done, on your terms, with you still holding the trigger. One of those you can trust with a project you care about. We built for the other reason entirely. If you want the fuller argument, here is why human-in-the-loop works.