From vibe coding to context engineering: 2025 in software development (www.technologyreview.com)

Thoughtworks’ latest Technology Radar frames 2025 as a pivot from “vibe coding” — a loose, prompt-driven style popularized by Andrej Karpathy — toward what the industry calls “context engineering.” The report flags growing antipatterns and model reliability issues as prompts ballooned and developers became complacent with raw AI-generated code. In response, teams are experimenting with knowledge priming and tooling (e.g., Claude Code, Augment Code) to supply the right context so outputs are more consistent, reduce rewrites, and raise productivity. Crucially, Thoughtworks notes that context isn’t just more data: for forward engineering, higher-level abstractions can widen the solution space and better leverage generative creativity. The rise of agentic systems heightens the need for disciplined context management. Practical techniques include anchoring agents to a reference application, using teams of specialized coding agents to split contextual load, and leveraging emerging context platforms (agents.md, Context7, Mem0). Standards are starting to converge around the Model Context Protocol for linking LLMs/agents to context sources and the agent2agent (A2A) protocol for inter-agent interaction. The implication for AI/ML practitioners is clear: progress now depends less on scale and latency and more on engineering practices, protocols, and human-led design of contextual ground truths — placing software engineers, not just models, at the center of reliable AI-enabled development.