# New in LangSmith: 4 Updates for Managing Agents in Production

## Executive summary

LangSmith has introduced four major updates to manage AI agents in production, shifting the core unit of work from traces to the agent itself. Key features include LangSmith for coding agents, which ties AI spend directly to the resulting pull request (PR) for cost analysis; new sensitive data access controls for granular security; and LangSmith Tool Gateway, which centralizes tool access and handles authentication, ensuring agents never handle tokens.

## Key takeaways

- Agent-Centric Management: LangSmith is now organized around agents, allowing teams to manage traces, evaluations, and governance workflows for each agent across its full lifecycle, supporting different frameworks and runtimes (0:23).
- AI Spend Visibility (Coding Agents): LangSmith for coding agents links AI spend to the specific pull request produced. This allows teams to measure the cost of a merged PR and investigate cost drivers (model, skill, or runaway thread) (0:56).
- Secure Data Access Controls: New controls allow engineers to use trace metadata for debugging while restricting access to sensitive inputs and outputs, maintaining least privilege access for security teams (1:39).
- Centralized Tool Access (Tool Gateway): LangSmith Tool Gateway allows agents or MCP clients to access tools via a single URL. LangSmith handles credential attachment and logging on every call, meaning agents never touch a token (2:20).

## Technical details

- LangSmith for Coding Agents: This feature is agent agnostic and allows teams to track cost per merged PR, enabling investigation into what drives cost (e.g., model, skill, or runaway thread) (0:56).
- Sensitive Data Controls: Permissions can be set by user group and based on trace metadata, enabling engineers to debug with trace metadata while limiting access to inputs and outputs (1:39).
- LangSmith Tool Gateway Mechanism: Tools are registered as MCP servers, grouped into a kit, and provided a single URL. LangSmith verifies the caller and attaches the correct user/agent profile credentials on every call (2:20).

## Practical implications

- Improved governance and consistency when managing a large number of production AI agents.
- Direct financial accountability for AI usage, allowing teams to measure the ROI of coding agents against shipped PRs.
- Enhanced security posture by decoupling agent execution from sensitive authentication tokens via Tool Gateway.
- Better collaboration between security and engineering teams through granular data access controls.

## Topics

AI Agents, MLOps, LangSmith, Security, Cost Management, Why we rebuilt LangSmith around the agent, LangSmith for Coding Agents, Coding agent docs

Source: https://www.youtube.com/watch?v=G-4VnMltDlc
