# Software Engineering Is Becoming Plan and Review Louis Knight-Webb, Vibe Kanban | AI Engineer Europe 2026 | 20:23 Source: https://www.youtube.com/watch?v=W76woOYHlvY Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/software-engineering-is-becoming-plan-and-review Published: 2026-05-02 Tags: agents, coding-agents, developer-experience, planning, workflows ## TL;DR - AI tools reduce the time engineers spend writing code, but that time moves into planning, reviewing, testing, and shepherding changes to deployment. - Detailed planning usually saves review time, while front-end feature work often benefits from staying closely involved with the coding agent. - When agents take longer than five minutes, engineers need interfaces for managing several concurrent work streams without constant context switching. ## Summary Louis Knight-Webb argues that AI is moving software engineering away from typing code and toward planning and reviewing AI-generated work. He compares two working styles: plan heavily before starting an agent, or provide a loose request and spend more time correcting the result. The right balance depends on the task. Front-end feature work has too many interaction and state edge cases for complete up-front specification, while migrations, refactoring, and much back-end work can be planned and tested in advance. As agents run for five minutes or longer, engineers need to manage several tasks in parallel. Knight-Webb shows Vibe Kanban, a tool built for that workflow, then announces that the company is shutting down. He says the project will continue non-commercially, while the team moves on. He attributes the business failure to selling to individuals and startups rather than enterprise customers or token usage. ## Key ideas ### AI removes coding time by moving it into planning and review [01:45](https://www.youtube.com/watch?v=W76woOYHlvY&t=105s) Knight-Webb describes software work as planning, writing code, reviewing code, and reviewing other people's code. GitHub Copilot, ChatGPT, Cursor, and Claude Code have reduced the writing portion. That does not give engineers the same number of hours back. The displaced time moves into planning and reviewing AI output. He estimates that an engineer might get 20 minutes back for every 30 minutes previously spent coding, while the rest becomes planning and review. His claim is that AI increases output, but changes what engineers spend their day doing. ### Detailed plans trade more preparation for fewer review cycles [03:43](https://www.youtube.com/watch?v=W76woOYHlvY&t=223s) The plan-based approach uses a comprehensive plan document or a specification framework before an agent starts. Knight-Webb describes systems that repeatedly question the user until the model has exhausted its questions about the task. This up-front work gives the agent more information and removes edge cases before execution. The expected result is better output and fewer review rounds. The alternative is to give the agent a loose request, such as adding a contact form, then correct the result through several review cycles. He prefers planning when human time is the scarce resource because repeated back-and-forth with an agent is expensive. ### The best planning and review balance depends on the kind of work [06:04](https://www.youtube.com/watch?v=W76woOYHlvY&t=364s) Knight-Webb separates front-end feature development from back-end features, refactoring, migrations, and maintenance. Front-end features are difficult to specify completely because they contain state, interactions, animations, styles, and functionality. He prefers staying in the loop with an agent for this work. Back-end features can often use a test-driven approach, while refactoring and migrations should be planned heavily and allowed to run without constant human involvement. He condenses the advice into a time trade-off: five minutes of planning can save 30 minutes of reviewing AI-generated code. ### Longer agent runs make waiting a workflow problem [07:27](https://www.youtube.com/watch?v=W76woOYHlvY&t=447s) As models improve and agents gain more tools, each execution takes longer. Knight-Webb contrasts GitHub Copilot completing a line in seconds, the original Cursor completing a file in about 30 seconds, and Claude Code runs that can take five or ten minutes. Agents now run type checkers and tests, and Playwright MCP can take longer still. He accepts the wait when it improves accuracy. He expects agents to eventually run front-end projects, click through them, find bugs, and make corrections themselves. That would remove much of the current human back-and-forth. ### The five-minute threshold changes how engineers spend attention [09:54](https://www.youtube.com/watch?v=W76woOYHlvY&t=594s) Knight-Webb says five minutes is roughly the point where a person stops watching an agent's logs and starts doing something else. If runs take 20 minutes, an engineer cannot simply wait at the terminal. The response is to run several agents at once. When one task is being reviewed, another can finish, keeping the engineer moving between completed pieces of work. This changes software development from deep focus on one task into managing several streams of agent work. ### Agent interfaces should protect human focus while supporting parallel work [12:00](https://www.youtube.com/watch?v=W76woOYHlvY&t=720s) Knight-Webb wants coding tools to assume that one engineer manages multiple streams of work. He calls the goal "focus maxing" and says constant movement between contexts can wear people down. A better interface would let an agent run for as long as possible before returning control to the human. It would help write tasks and plans, perform quality checks, support code review, and shepherd changes through deployment. He also expects humans to keep reading code before shipping, even when AI performs much of the review. ### Vibe Kanban is shutting down because its customer economics did not work [13:59](https://www.youtube.com/watch?v=W76woOYHlvY&t=839s) Knight-Webb announces that he decided to shut down Vibe Kanban shortly before the talk. He says the project had 30,000 monthly active users and 25,000 GitHub stars, and will continue non-commercially. The company had a subscription that helped people run agents such as Codex, but users might pay Vibe Kanban $30 while using it to spend $3,000 with an agent provider. Knight-Webb says companies making money in the market were selling to enterprise customers or reselling tokens. Vibe Kanban did neither, and he did not want to move into enterprise sales to compete for a late position in a mature market. ### The startup experience left Knight-Webb positive about trying again [18:13](https://www.youtube.com/watch?v=W76woOYHlvY&t=1093s) In the questions, Knight-Webb says he plans to take some time off and start another company, while his co-founder joins a lab and much of the team moves to new jobs. He would not change the decision to build Vibe Kanban and says the work increased his value as a human. His advice from running the company is to work with great people and learn what hard work looks like in practice. Asked what he would change, he gives a specific answer: he would hire someone who was very good at selling to enterprise customers. ## Notable quotes - "The answer, of course, is no. It has displaced work." (03:04) - "Spending 5 minutes of planning saves you 30 minutes of reviewing AI-generated code." (06:57) - "When we cross that 5-minute mark, you have to change your behavior." (09:54) - "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 no way to live." (12:30) - "It's actually very difficult to make money in the current environment." (16:06) ## Tools & references mentioned - GitHub Copilot - ChatGPT - Cursor - Claude Code - Codex - Playwright MCP - Vibe Kanban - OpenAI - Cloudflare - Agent Labs - SWE-bench Verified - GitHub ## Who should watch - You are deciding how much planning to give coding agents before they start, especially for front-end work or migrations. - Your agents run long enough that waiting at one terminal is becoming a bottleneck, and you need to coordinate parallel tasks. - You are building an AI developer tool and want to understand the customer and token economics Knight-Webb encountered. ## Related talks - [Developer Experience in the Age of AI Coding Agents](https://aietalks.com/talks/developer-experience-in-the-age-of-ai-coding-agents) (Max Kanat-Alexander, Capital One, 18:20) - [Don't get one-shotted: Use AI to test, review, merge, and deploy code](https://aietalks.com/talks/dont-get-one-shotted-use-ai-to-test-review-merge-and-deploy-code) (Tomas Reimers, Graphite, 05:45) - [How to Kill the Code Review](https://aietalks.com/talks/how-to-kill-the-code-review) (Ankit Jain, Aviator, 16:26) - [Everything We Knew About Software Has Changed](https://aietalks.com/talks/everything-we-knew-about-software-has-changed) (Theo Browne, @t3dotgg, 16:02) - [Software Development Agents: What Works and What Doesn't](https://aietalks.com/talks/software-development-agents-what-works-and-what-doesnt) (Robert Brennan, OpenHands, 16:46)