Claude Code
Claude Code is a recurring example of an agentic development environment whose effectiveness depends on the surrounding harness rather than on model output alone. The corpus concentrates on context management, configurable effort, delegated subagents, skills, hooks, MCP gateways, and permission boundaries.
Patterns surfaced in the corpus
- Tool and context design evolve with model behavior: structured questions, task delegation, progressive disclosure, and agent-accessible documentation can replace heavier up-front prompting. [src]
- Practitioners describe repository instructions, document-and-clear workflows, commit-time gates, skills, and narrow MCP gateways as controls around delegated work. [src]
- Model selection controls capability while effort governs how much checking, reading, and tool use a turn receives; auto mode adds a permission-control layer rather than simply bypassing prompts. [src] [src]
Coverage expansion
Simon Willison reports advice from a Claude Code team fireside chat to let the main agent exercise judgment rather than hard-code every operational rule, while delegating routine coding subtasks to lower-power models when appropriate. The account is a report of advice and one user's project-memory workflow, not an official capability guarantee or a universal routing policy. [src]
For the workspace controls that make this kind of delegation inspectable, see Version Control And Agent Workflows.