LangSmith CLI Arrives
LangChain is shipping a new LangSmith CLI designed from the ground up with AI agents in mind. The tool gives both developers and coding agents the ability to interact with LangSmith entirely through the terminal, including fetching traces, curating datasets, and running experiments. Installation is straightforward via a single curl script.
What Are Skills?
"Skills" represent a novel approach to improving agent performance: curated instructions, scripts, and resources that agents can dynamically retrieve only when relevant to their current task. This "progressive disclosure" model prevents performance degradation that can occur when agents have access to too many tools simultaneously. Skills are portable and shareable as markdown files and scripts, making them distributable across different coding agents.
Three Core Skills Shipped
The initial LangSmith skills release includes three capabilities aligned with LangSmith's core workflows:
- Trace: Add tracing to existing code and query traces for debugging
- Dataset: Build and curate test datasets from examples
- Evaluator: Evaluate agent performance against those datasets
Impressive Performance Gains
Testing against Claude Code (Sonnet 4.6) showed dramatic improvements: agents without skills achieved a 17% pass rate on LangSmith tasks, while the same agents equipped with skills reached 92%. This demonstrates the potential for skills to meaningfully enhance agent capabilities in specialized domains.
Agent Development Loop
The skills enable a powerful iterative cycle for agent development: agents can add tracing to their own code, generate and analyze traces for debugging, create systematic test datasets from that data, build evaluators to validate correctness, and iterate on architecture based on results. This agent-driven improvement loop is positioned as critical infrastructure for the future of AI agent development.