Luisa Crawford
Mar 04, 2026 19:39
LangChain releases new CLI tools and skills system that dramatically improves AI coding agents’ ability to work with LangSmith ecosystem for tracing and evaluation.
LangChain has released a new CLI and skills system that boosted Claude Code’s performance on LangSmith-related tasks from 17% to 92%, according to internal benchmarks shared March 4, 2026. The tools aim to create what the company calls a “virtuous cycle” where AI agents can debug, test, and improve other AI agents.
The release builds on LangSmith Fetch, the CLI tool LangChain launched in December 2025 that brought trace access directly into terminals and IDEs. That earlier tool already demonstrated significant efficiency gains—up to 96% context savings compared to traditional debugging methods for large traces.
What Skills Actually Do
Skills are essentially instruction sets that coding agents load dynamically when needed. Think of them as specialized knowledge packs. The key innovation here is progressive disclosure—agents only pull in relevant skills for their current task rather than loading everything upfront.
This matters because previous research from LangChain showed that overloading agents with too many tools actually degrades their performance. By keeping skills modular and on-demand, agents stay focused.
The initial release includes three skill categories:
Trace: Add tracing to existing code and query execution dataDataset: Build example sets for testingEvaluator: Run agents against those datasets and measure correctness
The Agent Development Loop
LangChain is positioning this as infrastructure for agents that improve other agents. The workflow looks like this: a coding agent adds tracing logic to your project, generates traces during execution, uses those traces to build test datasets, creates evaluators to validate behavior, then iterates based on results.
Whether you buy into the “agents improving agents” vision or not, the practical value is clear. Developers working with LangSmith now have command-line tools that their AI assistants can actually use effectively. Installation runs through a simple curl script or npm package.
Performance Claims Need Context
The 17% to 92% improvement sounds dramatic, but it’s measuring a narrow benchmark—specifically how well Claude Code handles LangSmith-specific tasks without versus with the skills loaded. LangChain says they plan to open source the testing benchmark, which will let the community verify these numbers independently.
The underlying CLI boasts sub-100ms startup times through lazy loading, and supports multiple output formats including JSON for scripting and formatted tables for human readability. It can interact with projects, runs, datasets, examples, prompts, and threads within LangSmith.
LangChain also released a parallel set of skills for their open source libraries—LangChain, LangGraph, and the newer DeepAgents framework. For teams already embedded in the LangChain ecosystem, these tools should reduce friction significantly. For everyone else, it’s another data point in the race to make AI coding assistants actually useful for specialized development workflows.
Image source: Shutterstock
