After Sonnet
- 5 came out, both Anthropic and users found that its coding capabilities were particularly strong, and an adventure based on coding capabilities began. The concept of Ralph Loop is very simple: continuously iterate the interaction and feedback between the input and the large model in a non-stop loop, and constantly repeat this process. During this process, follow up the plan status and correct errors... ClaudeCode is not a typical Ralph Loop. It places more emphasis on progressive engineering, state recording and hook mechanisms, and HITL (human in the loop) to ensure a complete agent experience. This is one of the reasons why ClaudeCode was so popular when it was launched in February 2025 - it can really work stably based on reliable model coding capabilities. On this basis, the final architecture is shown in Figure
- With each iteration of the model, ClaudeCode is also strengthening its harness, absorbing external and internal best practices, MCP, Subagent, Skill, Goal... The more it is used, the faster the progress of the architecture will be. When we look back at the model progress and the evolution of Agent and Harness over the past year or so, they have all evolved step by step, although the speed of this evolution seems to exceed the speed of the past 10 years (my personal feeling). The saying that true knowledge comes from practice has not changed, but it has been greatly accelerated. As the request limit is reached again and again, the model, the agent and the people involved are all evolving, and the bridge in the middle is data. Everything is still evolving. . .
