Software Engineering Is Becoming Plan and Review — Louis Knight-Webb, Vibe Kanban
TL;DR
Software engineering is shifting from writing code to planning and review — Louis Knight-Webb argues tools from GitHub Copilot to Cursor to Claude Code have steadily shrunk the “write code” slice of the job and expanded the time spent specifying work up front and checking AI outputs afterward.
His core heuristic is blunt: 5 minutes of planning can save 30 minutes of review — he says detailed specs, exhaustive clarifying questions, and plan-heavy workflows usually beat “YOLO” prompting because bouncing back and forth with half-finished AI work is what really burns human time.
Front-end feature work is still the exception — Knight-Webb says backend features, migrations, and refactors can be heavily planned or even test-driven with little human-in-the-loop work, but stateful front-end development still resists full specification because of interactions, styles, animations, and edge cases.
Longer-running agents will force a new operating model for developers — as coding agents move from seconds-long completions to 5-10 minute runs with type-checking, tests, and Playwright-style QA, engineers will increasingly manage parallel streams of work instead of staring at one terminal.
He thinks browser-based self-QA is the next big breakthrough — demos with Chrome or Playwright MCP already hint at agents that can click around, find bugs, and fix them themselves, which could eliminate much of today’s tedious back-and-forth review loop.
The talk ends with a live shutdown of his own startup, Vibe Kanban — despite 30,000 monthly active users and 25,000 GitHub stars, he says the business was stuck between enterprise sales and token resale economics, with customers paying Vibe Kanban $30 to help them spend $3,000 on Codex.
The Breakdown
The job is becoming “plan and review”
Louis Knight-Webb opens with the real question behind the title: what are software engineers going to do all day once AI gets really, really good? His answer is that the old mix of planning, coding, and reviewing is tilting hard away from code-writing itself — first with GitHub Copilot, then ChatGPT, then Cursor, then Claude Code — until the job becomes mostly planning tasks and reviewing outputs.
Two ways to work with coding agents
He lays out two clear modes: either invest heavily upfront in planning, or keep the spec loose and pay for it later in review. The first looks like long plan docs, spec frameworks, and the model interrogating you until it has no more questions; the second is more like “let’s add a contact form” and then cleaning up the mess in rounds. His punchline is memorable and practical: spending 5 minutes planning saves 30 minutes reviewing AI-generated code.
Where spec-heavy workflows work — and where they don’t
Knight-Webb says not all work behaves the same. Backend feature development, migrations, and refactors are good candidates for plan-heavy or test-driven workflows, where the human can stay mostly out of the loop; front-end feature work is different because it’s full of state, interactions, animations, styles, and weird edge cases that are hard to fully spec in advance. That’s the quadrant where he still wants to stay close to the agent.
Agents are running longer, and that changes the job
He traces a timeline from Copilot’s seconds-long line completions to Cursor’s roughly 30-second file edits to Claude Code now producing useful 5-10 minute runs. That’s happening because agents are doing more than just returning code — they’re running type checkers, tests, and increasingly heavier toolchains like Playwright MCP. His point is that longer runtime is actually good if it reduces human intervention, but once jobs stretch past about 5 minutes, you can’t just sit there watching logs anymore.
From single-task coding to managing parallel work
That 5-minute threshold is his behavioral breakpoint: beyond it, developers need a new interface and a new mindset. He describes the emerging pattern as running multiple agents in parallel so that while one task is executing, another is ready for review — less “lock in on one problem,” more managing several streams without frying your brain. He jokingly calls the goal “focus maxing,” meaning tools should preserve human context instead of yanking people in and out every 30 seconds.
What he built in Vibe Kanban
That theory led to Vibe Kanban, which he describes as a way to parallelize coding agents with multiple workspaces, code diffs, inline comments, previews, and support for tools like Codex and Claude Code. He notes, with some founder pride, that they were doing this as early as June 14, 2025. His wishlist for the ideal tool is very manager-shaped: help write tasks, help QA work, support code review, and shepherd changes all the way from finished task to deployed PR.
The plot twist: he shuts the company down on stage
Then the talk swerves. Knight-Webb reveals that he decided on Tuesday to shut Vibe Kanban down, so instead of pitching the product, he uses it live on stage to publish the shutdown post. The room laughs, groans, and applauds while he explains that despite 30,000 MAUs and 25,000 GitHub stars, the business didn’t work because the winners in this market are either selling to enterprise or reselling tokens, and Vibe Kanban was doing neither.
Why he’d still do it all again
In the Q&A, he sounds more relieved than devastated, saying the burden of staff, investors, and company-building had lifted. He says he wouldn’t do anything differently except maybe hire someone great at enterprise sales, and that the whole experience increased his value “as a human.” His practical founder lesson is simple: work with great people, learn what real hard work feels like, and know that “playing for eighth place” in a mature market is no fun.