LangChain launched Preview Builds, a new LangSmith Deployment feature (in Public Beta) that spins up temporary, production-like staging environments directly from pull request branches. Instead of creating a full deployment for every PR, teams can test agent changes—prompts, tools, models, dependencies, integrations—in an isolated environment linked to the parent deployment, then tear it down when review is complete. The feature automatically syncs previews with the latest commit on the PR branch, keeping the running version up-to-date as developers respond to feedback.
Previews solve a real friction point: agent behavior often becomes clear only when the agent runs in production-like conditions, and local testing rarely gives cross-functional teams a consistent place to review. With Preview Builds, product managers, domain experts, and QA can test the same agent version in the same environment without cloning repos or reproducing local setups. Collaborators can inspect traces, test tool calls, and validate failure paths. Teams control resource usage through idle TTL (auto-deletion after inactivity) and max concurrent previews limits.
The feature is available on LangSmith Cloud for deployments connected via the GitHub integration, with two trigger modes: 'Every PR' (useful when most changes need review) or 'Label only' (for selective previews). Secrets are scoped per preview and can be overridden for sensitive services, mitigating risk when external contributors open PRs.
For agent engineering teams, this reduces the friction between code review and deployment testing. The risk is modest integration overhead with GitHub CI/CD and secret management, but the UX win—shared staging linked to PRs—directly addresses a common bottleneck in agentic workflows. Watch whether this pattern spreads to other LLM platforms (Anthropic's Workbench, Vercel AI SDK, etc.).