Harness Engineering

All readers — why agent performance is determined by the harness around the model, not the model alone (Anthropic's core thesis on harness engineering)

Agent system authors — from augmented LLMs to five workflow patterns and the autonomous agent (Anthropic's taxonomy of agentic system patterns)

Agent system authors — making agents work across multiple context windows (the initializer/coder dual-agent harness pattern for multi-session autonomous tasks)

Agent system authors — GAN-inspired three-agent architecture (Planner / Generator / QA building complete applications together over multi-hour autonomous sessions)

Platform / infra engineers — decoupling the brain from the hands (Session as a persistent log, Sandbox as a disposable execution environment, interfaces that outlast any specific harness implementation)