LangChain brings pull-request cost attribution to LangSmith
LangChain added a LangSmith dashboard that associates coding-agent trace costs with pull requests while separating spend from sessions that produced no pull request.
LangChain has introduced pull-request cost attribution to LangSmith, giving engineering teams a dashboard that compares coding-agent spending with repository activity. The feature is available in public beta.
The dashboard separates costs associated with pull requests or merge requests from “unattributed” spending on sessions that did not lead to one. It offers an observability view of agent usage and delivery artifacts. But the association does not establish that an agent caused a pull request, improved code quality or developer productivity, or generated a financial return.
LangChain says LangSmith attributes spend in coding-agent traces to the pull requests each agent contributed to, with outcomes covering GitHub pull requests and GitLab merge requests. The company says the model covers both a single thread that opens several pull requests and a pull request completed across several threads.
The dashboard breaks spending down by pull-request state and links metrics back to the underlying threads. Other views show the most expensive pull requests, work type, skill usage and how cost changes with conversation length.
The launch materials list tracing integrations for Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code and Deep Agents Code. LangSmith for Coding Agents is available in public beta on LangSmith Cloud.
The spending figures depend on the data entering the system. LangSmith can calculate model cost from token counts and configured model prices, or accept costs supplied directly on a run. Its coding-agent metadata contract requires a stable thread identifier on every run, along with repository, repository-provider and active-branch fields when the runtime can expose them.
LangChain says unattributed sessions can still cover work such as code review, investigation or codebase exploration. The launch materials do not provide a controlled comparison with unassisted development, independent validation of attribution accuracy, or measurements of time saved, defects, review burden and downstream production performance.
More news

Databricks makes AI Index public for model spending and coding agents

Kiro Crew launches persistent bots for recurring development work

Google opens Antigravity teamwork tools to external researchers
