
LangSmith now puts a single number in front of finance for coding agents: cost per merged PR. This is a Kill List entry for that number as a headline, and a keep for the attribution under it.
The short version
- On November 18, 2026, Jacob Lee announced LangSmith for Coding Agents: agent spend attributed to the pull requests it produced, PRs grouped as open, closed, or merged, and cost per merged PR as "what a unit of shipped work costs."
- The post says merged PRs are a useful signal and not the only sign of productive work, and that sessions which never led to a PR are tracked as unattributed. Kill cost per merged PR as the headline.
- The count stops at merge, and three things sit past it: the reviewer's hours, whether the change stayed, and the investigation and review work that never opens a PR.
- I shipped 39 agents and 13 ended with zero references anywhere else.
- Report cost per change still standing at day 30, with review minutes in the cost, and give the unattributed row names and owners.
What shipped
I read LangChain's post through a fetched summary and a set of quotes pulled from it, so I'll stay close to what those returned.
The feature is in public beta on LangSmith Cloud. Spend in each trace is attributed to the PRs or merge requests the agent worked on. One thread can open several PRs and one PR can take several threads. PRs are grouped by state. The dashboard has spend by PR state, work-type labels such as features, bug fixes, investigation, and code review, the most expensive PRs, and a before-and-after comparison when the team changes models. The post cites a Menlo Ventures figure of $34B in enterprise spend on coding in 2026, and says the spend of LangChain's own coding agent, Robin, was one reason the team built a model router.
Two sentences carry the argument. "Cost per merged PR shows finance what a unit of shipped work costs." And, on the other side, merged PRs are "a useful signal" but "not the only sign of productive work," because agents also review and investigate, which often ends without a PR. That spend is tracked "separately as unattributed."
The attribution is the valuable part and I'd adopt it. Tokens by PR beats tokens by seat. The before-and-after on a model change is the right experiment to make easy.
Three things past the merge
The reviewer. The numerator is agent spend. A person read the diff, and that person is not in the ratio. In SaaStr Runs $20M on 3 Humans and 20 Agents. Revenue per Human Hides the 13-Hour Review Day. I quoted Jason Lemkin on a recent 13-hour day that went entirely to reviewing and fixing agent work. A switch to a cheaper model can lower cost per merged PR and lengthen the review. The dashboard would show that as a win. I did not find human review time in the post.
Whether it stayed. Merged describes one day. I built 39 agents in 80 days and shipped every one of them, and the audit in 39 PM AI Agents Deployed: What Stuck, What Died, and Why found 13 with zero references from anywhere else in the system. All 39 were published, which is my equivalent of merged. A cost-per-merged number would have averaged the 13 in with the 26 and looked fine. I did not find reverts or rework in the post either.
13 of 39
The work with no PR. The post is honest that investigation and review often end without one. But a ratio next to a row called unattributed tells a finance reader which one is the result and which one is the leak. The investigation that concluded "don't build this" produces no PR on purpose, and it may be the cheapest good decision of the month.
There is a scale note from LangChain's own data. In Label Before You Route I wrote up their router A/B across 973 threads, where 29.2% of routed threads ended in a merged PR. If that is typical, most threads are outside the numerator of this metric.
What to report instead
Cost per change still standing at day 30. Agent spend plus review minutes at a loaded rate, divided by merged PRs that were not reverted or substantially rewritten in 30 days. The PR-state table already has open, closed, and merged. This adds one state after merged.
The unattributed row, named. Investigation, review, and abandoned, each with a share of spend and an owner. Abandoned is the only one of the three that is waste by default.
Review minutes per merged PR, as its own line. So a model change that moves cost from tokens to people is visible the week it happens.
None of this needs a new tool. The first comes from joining the PR list to reverts. The second is a relabel. The third is a field reviewers fill or a timer on the review.
This week
Take last month's merged agent PRs and mark the ones that were reverted, rewritten, or are referenced by nothing. Divide the spend by the rest. That is the number to bring next to cost per merged PR. This sits in Enterprise AI Agents because it is the internal version of the outcome question every agent vendor is being asked.
Related answer: Is cost per merged PR a good metric for coding agents?
Sources: Jacob Lee, "LangSmith for Coding Agents: Now track spend by outcome," LangChain, November 18, 2026, read as a fetched summary plus targeted quotes: the attribution, the PR states, the three quoted phrases, the unattributed row, the supported agents, and the Menlo figure as the post cites it. I did not read the page line by line and did not open the LangSmith docs. The 39, 13, and 26 are mine, the 13-hour day is Jason Lemkin's, and the 29.2% across 973 threads is from Sydney Runkle and Eugene Yurtsev's October 1 post as I wrote it up, all in the posts linked above.
Frequently asked
What did LangChain ship for coding agents?+
On November 18, 2026, Jacob Lee announced LangSmith for Coding Agents in public beta on LangSmith Cloud. As the post describes it, spend in each trace is attributed to the pull requests or merge requests the agent contributed to, PRs are grouped as open, closed, or merged, and the dashboard reports cost per merged PR. It supports Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code, and Deep Agents. I read the post through a fetched summary and targeted quotes, not line by line.
What is wrong with cost per merged PR?+
The count stops at merge. The reviewer's time is outside the spend, a merged change may be reverted, rewritten, or used by nothing, and agent work that never opens a PR lands in a row called unattributed. The post itself says merged PRs are a useful signal and not the only sign of productive work.
Does the LangChain post address review time or reverts?+
Not that I found. The word review appears for agent code-review threads. Human review time and whether merged code stays are not mentioned in what I read.
What should replace it?+
Cost per change still standing at day 30: agent spend plus review minutes, divided by merged PRs that were not reverted or rewritten in the window. Next to it, the unattributed row split into investigation, review, and abandoned, each with an owner.
What is your own evidence?+
I built 39 agents in 80 days and shipped all of them. When I audited them, 13 had zero references from anywhere else in the system. Every one of the 39 was published, my equivalent of merged. A cost-per-merged count would have treated the 13 the same as the 26.
Is the attribution itself useful?+
Yes. Tying spend to a PR is better than a bill by seat or by engineer, and the before-and-after view on a model change is the right experiment. The objection is to making merged the unit that finance sees.


Sign in with LinkedIn to leave a comment.
Sign in with LinkedIn