QCon London AI Coding State of the Game: More Capable, More Expensive, More Dangerous Coding Agents

Birgitta Böckeler's QCon London keynote reviewed the evolving AI coding landscape, highlighting a shift from vibe coding to autonomous coding agents and agent swarms. Key developments include context engineering with granular skill loading, headless CLI modes connecting to CI/CD pipelines, and multi-agent orchestration via Claude Code's Agent Teams. She raised two major concerns: a worsening security landscape driven by prompt injection attacks (including a real incident where an unsupervised agent leaked secrets to an NPM registry), and rising costs of agent-based development. She proposed a risk framework based on mistake probability, impact, and detectability, and emphasized that security is a conceptual rather than purely technical problem. Her closing point: AI coding amplifies existing engineering practices, good or bad.