AI NewsDev toolsAnnouncement

LangSmith now ties coding agent spend to the merged pull request

LangSmith now links each coding agent run to the GitHub or GitLab pull request it produced. A team can see the dollar cost of a merged PR and the dollar cost of runs that led nowhere.

AI News

Editorial3 min read

LinkedInX
Signal-field composition representing cost reporting for coding agents.

Why it mattersMost teams running coding agents know the monthly bill and nothing else. A number that ties spend to the thing the team cares about, a merged PR, changes the conversation with finance.

LangChain's Jacob Lee posted a new LangSmith feature on 18 November 2026. The product now links each coding agent run to the GitHub pull request or GitLab merge request it produced. A team can see the dollar cost of a merged PR. The team can also see the dollar cost of every agent run that never led to one.

LangChain says the feature works with five coding agents at launch: Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code, and LangChain's own Deep Agents. The feature is in public beta on LangSmith Cloud. Coding agent traces count toward the regular LangSmith tracing usage. The post does not list a separate price.

What the dashboard shows

The LangSmith coding agent overview page groups spend by the PR the agent worked on. Each PR carries its underlying agent threads. A reader can click through from a cost line to the actual conversation. The page lists five cuts of the data:

  • Cost by PR state: open, closed or merged.
  • Cost by work type: feature, bug fix, investigation or code review.
  • Skills the agents called, with the spend each one drove.
  • The most expensive PRs in the period.
  • How cost scales with conversation length.

Unattributed spend is tracked on its own line. That is the spend from agent sessions that did not open a PR. LangChain's own words in the post: "Merged pull requests are a useful signal, but they're not the only sign of productive work."

The business case LangChain states

LangChain states one market figure in the post: "Coding is the largest category of enterprise AI application spend, reaching $34B in 2026." The number is attributed to LangChain. The post does not say which report produced it.

The product idea builds on an earlier LangChain piece published on 17 November. That piece set out the agent as the top-level record in LangSmith. In it, LangChain cited an engineering lead whose coding agent bill grew six times in two quarters. LangChain says this new feature is the second half of that work. Once the agent is a named record, its cost needs a named outcome.

What the post does not claim

LangChain names no customer logo and reports no measured ROI. The post lists work type categories as a feature of the dashboard. The five categories are written by the agent itself from the thread content. The categorisation is a model guess. A team should expect the usual "unknown" or "miscategorised" rows at launch.

The attribution also depends on an agent opening a PR. Some agent runs will not produce one. An agent might review a human's PR. It might write a one-off script, or debug a failing test in place of a user. Those runs show up in the unattributed line until the model learns to tag the work differently.

What to do this week

The feature is live in public beta on LangSmith Cloud. A team already using LangSmith for a coding agent trace can turn the view on today. The first number arrives the same week. The honest first test is this: pull one week of traces and open the PR cost chart. Then check whether the five work types match what the team believes it was doing. That gap, more than the dollar figure, is the first read on whether the dashboard is working.

Source

SourceLangChain

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project