Agent design, skills and context engineering

Agent architecture is moving from tool-search to structured skills and context engineering. - Box CEO Levie: context engineering is new moat - Failures stem from missing background, not model logic - Build retrieval, ranking, prompts and state optimization - 'Tool search' is unreliable at scale - Skills encapsulate APIs, usage patterns and prompts - Skills load on demand via classifier/router - Armin Ronacher: native Plan Mode in AI IDEs is essential - Plan Mode decouples intent, preview and atomic execution - Agent Skills standardized as open format (SKILL.md) - Coding practices may shift to 'by agent, for agent' codebases - AMP experiments show letting agents decide naming boosts efficiency
Key insights
Context engineering becomes the primary system barrier
Experts argue that as model reasoning improves, failures increasingly arise from insufficient contextual information rather than raw model capability. Building an efficient information supply chain—retrieval, heuristic ranking, system prompts and state/space optimizations—is now central. Box CEO Levie framed 'context engineering' as the next technical moat for agent reliability.
Skills replace naive tool-search in agent design
The community is shifting away from embedding large tool lists in prompts toward modular 'Skills' that include APIs, usage guidance and best practices. Skills are loaded on demand via intent classifiers or routers, reducing token waste and improving reliability. An open Agent Skills format with SKILL.md frontmatter has been promoted and adopted across multiple agent vendors.
Plan Mode in AI IDEs separates planning from execution
Armin Ronacher advocates native 'Plan Mode' as a core IDE feature: agents must first produce a detailed plan indicating files to read or modify and steps to execute, allowing human pre-approval and atomic execution. Plan Mode improves transparency, enables step-wise permissioning, and reduces unexpected large diffs during agent edits.
Codebases may evolve to favor agent-friendly conventions
Experiments show that when humans stop renaming or reorganizing AI-generated code and instead let agents define names and structure, agent efficiency increases. This suggests a potential emergence of codebases optimized 'by an agent, for an agent' where consistency with agent probabilistic expectations yields faster agent comprehension and modification.