Score your agent's context against five quality criteria, not just prompt wording

arXiv 2603.09619·medium signal

The arXiv paper 'Context Engineering: From Prompts to Corporate Multi-Agent Architecture' (2603.09619) reframes context as the agent's operating system and proposes five production-grade quality criteria to audit every context window: relevance, sufficiency, isolation, economy, and provenance. It layers intent engineering (encoding org goals/trade-offs) and specification engineering (machine-readable policy corpus) on top for multi-agent systems. Builder move: turn these five into a checklist or linter for what you stuff into a context window — most agent failures are a sufficiency or isolation problem, not a phrasing problem.