Anthropic Open-Sources Production Agent Playbook

- The playbook emphasizes a shift from "prompt engineering" to "context engineering," which treats the model's limited attention as a resource to be managed by curating the entire set of information it receives, including tool definitions, message history, and retrieved documents. - A core architectural pattern detailed is "progressive disclosure," where an agent initially loads only the name and a brief description of a "skill" (a reusable workflow), and only loads the full instructions on-demand, minimizing token usage. - For multi-agent systems, the guide details an "orchestrator-worker" pattern, which Anthropic uses for its own Research feature where a lead agent delegates tasks to specialized sub-agents that work in parallel. While this can boost performance by up to 90% on certain tasks, it can also use 15 times more tokens than a standard chat interaction. - Anthropic's approach to agent memory involves a client-side, file-based system where agents use tools to create, read, and update files in a dedicated `/memory` directory, giving developers full control over how and where memory is stored. - The company uses these same agentic principles internally for its Claude Code product, where AI agents now handle the majority of code reviews, replacing line-by-line manual review by human engineers. - The MIT license was chosen for the playbook, allowing developers to freely use, modify, and integrate the techniques into proprietary, commercial software with the only major requirement being the inclusion of the original copyright and license notice. - The guide cautions against over-reliance on complex frameworks, suggesting that developers should start by using LLM APIs directly, as many agentic patterns can be implemented with just a few lines of code, avoiding unnecessary layers of abstraction.