Anthropic Engineering Blog: Two-Agent Harness Architecture for Long-Running Autonomous Coding

Anthropic·high signal

Anthropic published a detailed engineering blog post on how they use a dual-agent harness — an initializer agent that sets up the project environment (init scripts, feature list, progress log) and a coding agent that works incrementally across multiple context windows. The system uses a claude-progress.txt file as a memory bridge between sessions and expands prompts into 200+ discrete end-to-end features tracked in JSON, with Puppeteer-based browser testing for verification. The post generated 3,460 likes and 435K views, spawning an emerging 'harness engineering' discipline across the industry.