Meta's acquisition of Manus for over $2 billion and Claude Code reaching a $1 billion run rate highlight the rapid advancements in agent technology. The article examines key design patterns in autonomous agents, reflecting on how they evolve to handle context more effectively. A significant challenge is context rot, where models struggle as context grows. Treating context as a limited resource, engineers like @karpathy emphasize the need for context engineering—strategically filling the context window with necessary information for optimal performance. One prominent design pattern is giving agents access to a computer, which enhances their capabilities. Agents like Claude Code and Manus utilize a filesystem and shell for persistent context and executing tasks. However, the article notes that while tools can empower agents, they can also overwhelm them if too many are defined, leading to confusion and inefficiencies. Many popular agents operate with a minimal set of tools, using a hierarchy of actions that allow them to perform complex tasks without burdening their context windows excessively. Progressive disclosure of actions is another key strategy, where agents only unveil necessary information as required. This method streamlines context management and prevents information overload. Offloading context to the filesystem helps agents maintain efficiency over long tasks, allowing them to read back essential information when needed. Caching context is vital for cost management, as agents with effective caching can operate more economically. Lastly, isolating context for sub-agents facilitates parallel processing and enables better task management in long-running operations. These design patterns collectively enhance the functionality and efficiency of autonomous agents, pushing the boundaries of what they can achieve.