Tony GorbProduct Designer & Builder

I’d love to join a product team for the long run and help build what comes next.

Optimind closed in October 2026.


View CV
LinkedIn (opens in a new tab)GitHub (opens in a new tab)

Designing human-in-the-loop decisions for AI coding agents

My coding agents kept leaving questions only I could answer, until I was buried in them. Making the list readable didn't help. Making each question answerable in one click did.

In this story3 sections
  1. Readable wasn't enough
  2. One question, one click
  3. The rhythm now

One morning in November I opened the page where my agents keep their open decisions, and I scrolled. And kept scrolling. I build my design tool with a team of AI coding agents, and whenever one of them hits a question only I can answer, it writes the question down, marks it for me, and carries on. Sensible. Polite, even. But nobody had been counting, and I hadn't been answering.

There were 68 of them. Some were real product calls: what agents should be called, who sees an agent's reply in a comment thread, whether a shared link should ask you to sign in first. Others were housekeeping. They all looked the same: a careful paragraph in the agent's own words, written for other agents.

Readable wasn't enough

My first fix was the designer's reflex. Make it readable. That same morning the page was rebuilt with every decision in full, the ones waiting for me on top, newest first, each point on its own line with room to breathe. It looked better. It was still a wall. Nicer type doesn't change the job, and the job was this: read a small essay, work out what's being asked, form an opinion, then go and tell an agent. Again and again. I typed something close to "woah, there are many many many open decisions" and asked for something different.

What I asked for was a question I could answer with one click, without anyone guessing on my behalf.

One question, one click

Each decision became a card. On top, one plain sentence of what was decided and one line on what it means for a person using the tool. The original text stayed underneath, behind a Details fold, because a plain-words summary can drift from what was actually agreed, and I wanted the source one tap away. Then the answers. A: confirm as written, with a one-line reason. B: change it, with a box for my own words. C: decide later. Extra letters only where the decision itself named a real alternative. Nothing invented. And where no plan backed a recommendation, the card said so instead of making one up. Cards that needed my judgement, things like privacy, names and what people lose, came first. Housekeeping sank to the bottom.

It didn't work the first time. I clicked an answer, watched "Saving…" blink, and saw my pick vanish. The page redrew itself on every save and wiped the choice. A small bug, but it taught me something: if an answer doesn't visibly stick, you stop trusting the whole form. Once it stuck, I went through the deck in one sitting and approved 65. The rest were changes, and those turned out to be the interesting part.

A B isn't a no. When I wrote that an export should follow the size of the selected frame, or that one agent per person felt wrong for a product built for teams, the agents didn't just log my note. They came back with new cards: the real options, what each would cost, and which one they'd pick. My half-sentence became a proper brief.

The rhythm now

By the end of the next week it had a rhythm. Agents card a question the moment they hit it. I answer in batches. On a good day my whole reply is two words, "Approve all", and when it isn't, my B goes off to become its own research and its own redesign. The agents keep moving, and I spend my attention on the few calls that actually need me.

Open decisions turned into cards
68
Answer options across those cards
208
Approved in the first sitting
65 of 68
Answers that confirmed as written
443 of 460

If you're designing an AI product that needs a person in the loop, don't make that person read the agent's notes. Hand them the question, the answer you'd pick and why, the source one tap away, and a way to say no in their own words. Then make sure the click lands somewhere the agents will read it.

Tony Gorb, Product Designer & Builder

Questions about this story, or a design role to talk about? Write to bytonygorb@gmail.com or find me on LinkedIn.

More stories