How to Build Agents That Improve Over Time: Model, Harness, Context | Interrupt London
LangChain frames agent improvement across model, harness, and context with new tooling for each layer
“Engine itself is nothing more than a deep agent in a sandbox with LangSmith CLI and some specialized subagents, some specialized skills, and some specialized prompts.”
LangChain presented a three-part framework for agent construction — model, harness, and context — and outlined how each layer can be improved over time through manual, human-in-the-loop, or fully automated means. They announced two concrete tools: smithtune (a CLI for supervised fine-tuning via LangSmith traces) and LangSmith Engine (an autonomous agent that monitors traces, clusters issues, and proposes harness fixes via PRs). Fully automated self-improvement is positioned as a long-term North Star, with current tooling still requiring human approval for harness-level changes.