Agents propose, humans approve
Notes from partway through building a hybrid team, where agents draft Shopify work and humans decide what ships.
My team runs as a hybrid of agents and humans. Agents draft Shopify work, route marketing requests, and stage QA for human review. People decide what ships. The rule that holds it together is short enough to say in a hallway. Agents propose, humans approve.
I want to be honest about where we are. This is a transformation in progress. I don’t have outcome numbers I’d put my name to yet, and I’d rather say so plainly than dress early impressions up as results. What I can share is how we started, why the review gate carries the weight, and what changes for the people doing the work.
We started with the work nobody loved doing, and adoption followed the relief. Our first agent didn’t write strategy and didn’t touch anything customer-facing. It took inbound marketing requests, asked the clarifying questions a coordinator would ask, and routed each one to the right person with context attached. Nobody mourned handing that off. The agents that followed draft first passes of routine Shopify changes and stage QA checklists before a human opens the build. Each one removed a chore, and interest in the next agent grew because the last one made somebody’s week lighter.
The review gate is what makes the speed safe. Every draft an agent produces stops at a person who can approve it, edit it, or discard it. Accountability stays where it always was, with a named human who understands the business. The gate earns its keep a second way too. Every edit and every rejection shows us where an agent’s instructions are thin, so the same checkpoint that protects quality also feeds the improvement loop. We tune prompts and process from real reviews instead of guessing.
For the team, the change shows up in the shape of a day. The humans didn’t go away. The waiting did. A request that used to sit in a queue until someone freed up now arrives as a reviewable draft, often the same day. People spend less time producing first passes and more time editing, deciding, and handling the cases that genuinely need judgment. Review turns out to be a skill of its own. A good reviewer reads an agent’s draft the way a senior developer reads a junior’s pull request, with care, and with the intent of making the next draft better.
If you’re thinking about starting, keep it small. Pick a chore rather than a showcase, something with a clear input, a clear output, and a person who already checks the result. Put the review gate in place before the first draft flows, so nobody has to trust the system on faith. After that, the relief does the recruiting for you. When an agent saves a teammate an afternoon, the next volunteer finds you.
I won’t predict where this lands. The measurement is being built alongside the operating model, and when the numbers are solid I’ll share the ones that hold up. What I can tell you today is why we’re doing it. The job I care about is closing the gap between the roadmap and the repo. Agents shorten that gap, and humans make sure what crosses it deserves to ship.