The module covers how ADEs are built and what that means for the human’s responsibility. The six-pillar architecture provides a framework for reading any ADE as an instance of a category: the bare LLM, tool augmentation, knowledge and memory, learning from experience, multi-agent coordination, and computer use.

The three ADE categories reflect different philosophies about transparency and control. Terminal and CLI-based agents expose the agent loop directly, give the developer maximum control over context and permissions, and compose naturally with other command-line tools. IDE-integrated agents embed AI in the environment the developer already uses, optimising for reduced friction at the cost of some visibility. Browser-based builders hide the architecture almost entirely, optimising for speed of creation. Each choice about transparency is also a choice about where the human’s responsibility is exercised.

Permission design and sandboxing are engineering decisions. What tools an agent has access to determines its blast radius, the scope of damage if it acts incorrectly. Minimal footprint is the principle: an agent should have only the permissions required for the specific task, prefer reversible actions over irreversible ones, and err toward doing less when uncertain.