Vibeleaderboard
← All Intel
Intel / article

LangSmith Preview Builds give agent PRs their own deployments

Source
langchain.com
Author
LangChain
Date

Context

Preview Builds give each pull request on an its own temporary, running copy of the agent, so reviewers can test how it behaves and not just read the code diff. LangChain has released the feature in Public Beta for LangSmith Deployment. The launch article explains the reason: a pull request can change prompts, tools, models, dependencies or integrations, and the effect of those changes "often becomes clear only when the agent runs." When a pull request triggers a preview, LangSmith deploys the branch in an isolated environment linked to the main (parent) deployment, without updating the parent. Each new commit pushed to the branch creates a new revision of the preview. Product managers, domain experts and QA can try prompts, exercise tool calls and inspect traces without cloning the repository or recreating the developer's local setup.

The demo video attached to LangChain's announcement post walks through the full cycle. It was reviewed through an automatic speech transcript, without analysis of the visuals. A developer adds a return-eligibility tool to a customer-support agent and opens a pull request, and a preview deployment is created automatically. The developer tests the preview, pushes a tweak that adds a days-remaining field, which produces a new preview revision, then merges. After the merge, the preview deployment is deleted, and continuous deployment builds a new revision of the main deployment. In the settings the presenter describes, previews can trigger on every pull request or only when a configured GitHub label is added. An idle time-to-live deletes inactive previews, about 72 hours by default in the demo. A concurrency cap limits how many run at once, and the demo leaves it at five.

The article raises one point about access. Previews copy the parent deployment's secrets when they are created and keep that initial set unless overridden. LangChain recommends credentials scoped to preview workloads rather than production for sensitive services, "especially if previews can be created from pull requests by external or less-trusted contributors." The beta covers deployments connected through the GitHub integration on LangSmith Cloud.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Read the source www.langchain.com
More from LangChain
Recommended reads
Comments

Checking sign-in…

Loading comments…