Tag
A breakdown of 12 reusable agentic harness design patterns observed in Claude Code, grouped into memory/context, workflow/orchestration, tools/permissions, and automation categories.
The author raises concerns about who controls data access for AI agents, questioning if there are established governance architectures, and shares a repo to explore the problem.
This paper catalogues five recurring MCP server architectural patterns observed across fifteen independently developed servers, providing a taxonomy with context, problem, solution, and consequences. It also documents anti-patterns, cross-cutting concerns, and quantitative evaluations including inter-rater reliability and transport overhead.
This paper introduces the concept of the stochastic-deterministic boundary (SDB) for production LLM agents and provides a methodology for selecting architectural patterns to improve reliability and performance.