Descriptions:
Louis Knight-Webb, founder of Vibe Canvas and organizer of the London AI Tinkers community, argues in this AI Engineer conference talk that software engineering is converging on a single core activity: planning what AI agents should do and reviewing what they produce. Drawing on his team’s experience reaching the SWE-Bench Verified leaderboard ahead of OpenAI, Knight-Webb traces how each generation of coding tools — from early GitHub Copilot to Cursor to Claude Code — has progressively compressed the time developers spend writing code and redistributed it toward planning and review.
The talk identifies two working styles that have emerged in response: a plan-heavy upfront approach that invests time in comprehensive specs and edge-case elimination before the agent runs, and an iterative YOLO approach that moves faster initially but requires more review cycles. Knight-Webb frames the choice as a time-distribution tradeoff rather than a quality one, and argues both converge on the same outcome as agent capabilities improve.
The second half focuses on the “time horizon” problem: as agent execution times grow from 30 seconds to 10-plus minutes, developers must shift from watching a single agent run to managing multiple concurrent workstreams. Knight-Webb introduces Vibe Kanban, a tool his team built to address exactly this parallelism challenge, and forecasts that autonomous frontend QA via tools like Playwright MCP will be the next major inflection point — enabling agents to close their own review loop without human intervention.
📺 Source: AI Engineer · Published May 02, 2026
🏷️ Format: Opinion Editorial
