Birgitta Böckeler, Distinguished Engineer at Thoughtworks, discusses the evolution of AI-assisted software development over the past year. Key shifts include the rise of Claude Code as the dominant coding agent, the move from MCP servers toward 'skills' and CLIs for more efficient context management, and the emergence of 'harness engineering' — a framework for building confidence in autonomous agents through feed-forward tools (coding conventions, architecture context) and feedback mechanisms (static analysis, test suites). She introduces a risk assessment model for deciding how much human supervision AI-generated code needs, based on probability of AI success, criticality of failure, and detectability of errors. Local models still lag behind for agentic use cases, and the field remains in a formative stage with many open questions around behavior validation, cost sustainability, and code privacy.
