Anthropic reveals three Claude agent patterns

Anthropic is advocating for "minimalist harnesses," betting that agents are more efficient when they orchestrate themselves rather than being managed by brittle heuristic layers. By moving orchestration logic into the model, agents can execute code to filter and pipe tool outputs internally. This "Dynamic Boundary Re-evaluation" treats the harness as "dead weight" to be pruned as model intelligence grows. Developers can reduce costs by up to 90% via prompt caching and shift from specialized tools to general-purpose environments like bash while maintaining typed tools for security.