Software Engineering Is Becoming Plan and Review — Louis Knight-Webb, Vibe Kanban

Louis Knight-Webb, founder of Vibe Kanban, presents a framework for understanding how AI is fundamentally reshaping the software engineering role from code-writing to planning and reviewing. Drawing on first-hand experience building a coding agent workflow tool — and achieving placement on the SWE-Bench Verified leaderboard ahead of OpenAI — Knight-Webb argues that the traditional ratio of engineering time has inverted: where engineers once spent the majority of their day writing code with little time on planning and review, AI has eaten the coding middle, and the remaining human work is now almost entirely plan and review.

Knight-Webb maps this shift across the evolution of AI coding tools. GitHub Copilot completed single lines in seconds. Early Cursor completed single files in roughly 30 seconds. Claude Code circa 2024 ran for one to two minutes; by 2026 Knight-Webb reports getting strong results from 5–10 minute agent executions. He predicts that within nine months, AI will begin reliably QA-ing front-end work using tools like Playwright MCP — a capability he calls the next major breakthrough — enabling models to run the project, click around, find bugs, and self-correct without human back-and-forth.

He identifies two primary working modes. The plan-heavy approach invests time up front — writing comprehensive markdown spec docs, using spec frameworks, or running models through exhaustive question loops — to reduce downstream review cycles. The review-heavy approach YOLOs a prompt and iterates through correction rounds. Knight-Webb's heuristic: five minutes of planning saves 30 minutes of reviewing AI-generated code. He notes that this trade-off varies by work type: front-end feature development is stateful and interaction-rich, making it difficult to fully pre-specify, while back-end feature development and refactoring/migration work are well-suited to test-driven, plan-first, no-human-in-the-loop execution.

A second structural shift concerns agent execution time crossing what he calls the "five-minute threshold" — the point at which an engineer can no longer sit and watch logs. Beyond this threshold, the job changes to managing multiple parallel agent streams. Knight-Webb calls this "focus maxing": designing workflows that allow humans to stay in deep focus while multiple agents run concurrently, yielding back to the human only when ready. He argues that future coding interfaces must be built around this manager-of-streams model, supporting task planning, QA of agent output, code review of AI-generated diffs, and automated PR shepherding (monitoring comments and reacting to CI automatically).

Vibe Kanban was built to address this parallelization problem. It provides a sidebar for running multiple isolated workspaces via Git worktrees, each running any of eight supported coding agents (including Codex and Claude Code). It surfaces diffs for review, supports GitHub-style inline commenting, and includes a live preview panel for visual QA. The tool had reached 30,000 monthly active users and 25,000 GitHub stars. However, Knight-Webb announces live on stage — having not yet told his investors — that Vibe Kanban is shutting down commercially. The core business model failed because the company was neither selling to enterprise nor reselling tokens; users paid $30/month to Vibe Kanban while spending $3,000/month with Codex. The project will continue as an open-source, non-commercial tool. He demonstrates the shutdown announcement live by using Vibe Kanban itself to generate and deploy a blog post to the company website, committing the change via an AI-generated pull request in real time on stage. In Q&A, his key lessons: hire someone excellent at enterprise sales earlier, and prioritize team quality above all else.