The Coding Job Is Now Planning and Review—And

Louis Knight-Webb stood on stage at an AI engineering conference and announced he was shutting down his startup. Live. Using the product itself to post the shutdown blog. The demo worked perfectly, which is maybe the most poetic way to explain why Vibe Kanban is closing: they built exactly what they set out to build, and it still wasn't enough.

But the real story isn't about a failed startup. It's about what Knight-Webb learned building tools for a job that's fundamentally changing shape.

The middle is disappearing

Knight-Webb's thesis is straightforward: AI is eating the actual coding part of coding. What's left is planning what you want the AI to do, and reviewing what it did. Everything else—the part where you're actually writing functions and debugging loops—is getting automated away faster than anyone expected.

He sketched out the evolution: Before GitHub Copilot, maybe you spent 60% of your day writing code, 20% planning, 20% reviewing. Then autocomplete tools arrived and writing code dropped to 40%. Then ChatGPT, then Cursor, then Claude. Now? "I actually am not really doing much code writing anymore," Knight-Webb said.

The natural assumption is that you'd get all that time back. You don't. The time just moves. "It has displaced work," he explained. "Time that was previously spent doing the coding has moved. It has moved to planning and reviewing."

This creates an interesting fork in how people work with AI coding tools.

Plan heavy vs. review heavy

Knight-Webb breaks it into two approaches, and which one you pick matters a lot for how productive you'll be.

The plan-based approach means spending significant time upfront writing comprehensive specs, maybe using frameworks that have the model interrogate you until it's exhausted all possible questions. The payoff: you spend way less time reviewing because you eliminated edge cases before the AI even started coding.

"Spending 5 minutes of planning saves you 30 minutes of reviewing AI-generated code," Knight-Webb said. That's the whole game in one sentence.

The review-heavy approach is what happens when you YOLO it. "Let's add a contact form to the webpage." Then you're going back and forth fixing styles, catching edge cases, iterating. It feels faster because you got started immediately, but you pay for it in review cycles.

Knight-Webb argues you should almost always choose plan-heavy if you value your time, with one major exception: front-end feature development. "It's basically impossible to kind of really spec everything out. There's so many edge cases and you know, front end is very stateful." For front-end work, being in the loop makes sense. For backend features, migrations, refactoring? Plan it properly and let the agent run.

But here's where it gets weird: as agents get better, they run longer. And that breaks something fundamental about how we work.

The five-minute threshold

GitHub Copilot completed a line in seconds. Early Cursor took 30 seconds for a file. Claude now runs for 5-10 minutes on a single task. Knight-Webb predicts that in nine months, AI will be QA-ing front-end work—actually running your app, clicking around, finding bugs—and that could push execution times past 20 minutes.

"I think 5 minutes is roughly the time when you can like sit there and wait for something, watch the logs, probably more realistically like browse Twitter," he said. Cross that threshold and you can't just wait anymore. You need a different workflow entirely.

The solution is parallelization—running multiple agents simultaneously so when you finish reviewing one piece of work, another is ready. This is what Vibe Kanban was built for: multiple workspaces, each running different coding agents, with integrated diff views and preview capabilities.

But this introduces a problem software engineers haven't really faced before: you're not deep-diving into one problem anymore. You're managing multiple streams of work, context-switching between different agents finishing different tasks. Knight-Webb calls this "focus maxing" (he's coining it, apparently), and it requires new interfaces.

"It should embrace the fact that you can't pull humans out of something and back into something else every 30 seconds cuz it just fries their brain and it's not it's no way to live," he explained. The ideal tool would help you write better task specs, QA the work, review code, and shepherd changes through deployment—all while protecting your ability to actually think.

Why Vibe Kanban shut down

Here's the uncomfortable part: Vibe Kanban had 30,000 monthly active users and 25,000 GitHub stars. The product worked. People loved it. And it's shutting down anyway.

The problem was economic, not technical. "Everybody who is making money is doing two things: they're selling to enterprise and they're reselling tokens," Knight-Webb said. Vibe Kanban did neither. They weren't a coding agent themselves—they were a workflow tool that helped you run agents like Cursor or Claude.

People would pay Vibe Kanban $30/month, then use that subscription to facilitate spending $3,000 with Claude. The unit economics don't work. Their users were individuals and startups, not enterprises. They could have pivoted upmarket, but Knight-Webb was blunt: "It's a mature market at this point and it's no fun playing for eighth place."

The project continues as open source, non-commercially. The team found good jobs at other AI labs. Knight-Webb seems... relieved? "I feel like a kind of weight has been lifted almost," he admitted when someone asked if he was sad.

What strikes me about this whole story is that Knight-Webb built exactly what his thesis said developers would need—tools for the plan-and-review workflow—and it still wasn't enough to build a sustainable business. Maybe that's because the problem is real but the willingness to pay for solutions is lagging. Maybe it's because this transition is happening so fast that the tooling market hasn't stabilized. Maybe it's because developers are stubborn about changing how they work.

Or maybe it's simpler: the people making money in AI right now are the ones selling the fundamental models or the enterprise wrapper services. Everything else is still being figured out, including whether there's actually a business in making the plan-and-review workflow less painful.

Knight-Webb wouldn't change anything, he said. "It was like the most interesting thing I've worked on." But he'd hire someone good at selling to enterprise next time.

The work changes before the business models catch up. And sometimes you build the right thing at the wrong altitude, solving a real problem for people who aren't ready to pay for the solution. That's the gap Vibe Kanban fell into.

—Yuki Okonkwo