blog.langchain.com
|
ksl
|
|
LangChain published a detailed framework for thinking about coding agents as two distinct halves – the model and the harness, where the harness covers everything from filesystem access and sandboxed execution to context compaction and cross-session memory. The piece introduces patterns like the Ralph Loop for long-horizon tasks and AGENTS.md as a memory standard, alongside their new deepagents library for building harness components. What stands out is the claim that optimizing harness design alone, without retraining the underlying model, produced meaningful benchmark gains. That finding tracks with what teams at Anthropic and OpenAI have been converging on: agent performance is increasingly an infrastructure problem, not just a model capability one.
