Track engineering output

See how your teams ship with AI in Work Intelligence, from AI-written code and pull request cost to review, quality, and tasks.

The Engineering view in Work Intelligence connects AI coding spend to the pull requests it went into, and shows how your teams ship: speed, review, quality, and the files and tasks behind them.

Who can see this: Admins see the whole organization and can filter by team. Managers see their reporting line, and every section except Skills. Members can't open Engineering.

Engineering opens on Overview. The other sections are tabs across the top: Pull Requests, Tasks, Hotspots, Languages, and Skills. Pick a range in the page header: Last 7 days, Last 30 days, Last 90 days, Last 365 days, or Month to date. One range applies to every section, and it stays selected when you switch tabs.

Most of Engineering is built from your GitHub pull requests. Until an admin connects GitHub, those parts say "Connect GitHub to see engineering metrics". See Connect GitHub. Admins can click Connect GitHub there. Managers see "Ask an admin to connect GitHub."

Overview

Overview is a grid of blocks. Hover a number with a dotted label to see what it counts.

BlockWhat it shows
AI workCode changes, Commits, Pull requests, Edits, and Edit acceptance from Claude Code; Cursor and Codex work isn't in these. Then Agent co-authored, the share of merged lines from commits that name a coding agent as co-author, and AI-assisted PRs, the share of merged pull requests with at least $1 of linked AI spend, such as "12 of 40 merged". These two cover every coding agent.
AI usage and costAgent hours (Claude Code agents only, as Claude reports it), 2+ agent sessions at once, Agent comments per PR, PR attributed cost, and, with Linear connected, Agent co-authored on bug fixes and AI cost per completed task.
Where the time goesHow long pull requests spend in each stage: Draft, Waiting for review, In review, and Approved to merged.
SpeedMerged, Time to merge, Waiting for review, and Open 7+ days, with a Pull request size chart.
ReviewMerged without review, Reviewed by bots only, Rubber-stamped, Review rounds, and Comments addressed, with a Review depth chart.
QualityRevert rate, Code turnover 30d, and Own rework 14d, with Turnover: agent co-authored vs other, which says whether agent co-authored code or other code was reworked more.
Merged per week, by who wrote itMerged pull requests each week, split into Mostly agent co-authored and Mostly other or not scored. Hover a week to see how many were AI-assisted.
Reviewer loadReviews split between People and Agents. Click either to see who gave them, and View all for each reviewer's reviews, comments, and response time.
Spend by effortSpend by reasoning effort level. Admins only. It always covers the whole organization; with a team picked, it says "Showing the whole organization. This can't be filtered by team yet."
Teams, People, AgentsThe five busiest of each by pull requests merged. Hover a row for time to merge, agent co-authored share, AI-assisted pull requests, and AI cost. View all lists everyone.

2+ agent sessions at once shows how many people ran more than one coding agent session at the same time, and the share of agent time spent that way. Cowork isn't counted. The line under it names the days it covers, which can be fewer than the range you picked. It shows how someone works, not how much they get done.

Agents are pull request authors that are GitHub Apps, such as coding agents that open pull requests themselves. In Teams, they're counted under Agents and bots. A pull request author Tokenize can't match to someone in your directory is listed by GitHub login, and counted under Unmapped GitHub logins in Teams.

When most of the coding-agent spend in view is in tools other than Claude Code, the Claude Code figures in AI work are replaced by one line, "Hidden: these count Claude Code only, and most coding-agent spend here is in other tools.", and Agent hours isn't shown. Agent co-authored, AI-assisted PRs, and the rest of the page stay, since they cover every coding agent.

On Last 90 days and Last 365 days, a large organization may see a note such as "Built Nov 18. Long ranges for large orgs are rebuilt each night." Those figures are as of that day. Pick a shorter range to see today's activity.

A stat shows "—" when Tokenize didn't receive that measurement. "Not measured yet" on turnover means no pull request in the range is old enough to compare.

Pull Requests

Lists pull requests with the AI spend attributed to them, from Claude Code, Codex, and Cursor sessions.

At the top, one stat per provider shows the cost that landed on pull requests and how many pull requests it landed on. A provider whose usage can't be read shows "—" rather than $0. Cost incomplete means some of that usage had no price, so the amount is a minimum. When some Cursor spend isn't on a pull request, a line under the stats says how much, and why, for example "Cursor conversations that edited no files (chat, reading, running commands)", "Cursor edits never committed", or "Cursor conversations from the last day or two, which may still be committed". Grok Bot spend reads "Grok Bot (no commits)" and is split by bot, named from Cursor's team audit log, for example "$400.00 “PR Reviewer” running Daily PR digest". Spend placed on a bot from its routine runs is marked "≈" and "estimated". Spend no bot can be matched to reads "unnamed bots". See See Cursor Grok Bots and How cost is calculated.

Cards are grouped under collapsible Open, Merged, and Closed headers. Pull requests Tokenize has no GitHub details for are grouped under No GitHub details and named by repository and number. Each card shows the pull request's tokens and attributed cost; hover the amount to see the cost by provider, Attribution, Unpriced usage, and Last AI activity. Each card opens the pull request on GitHub.

Attribution tells you how sure Tokenize is that the spend belongs to this pull request:

  • Direct link. The session was linked to the pull request directly.
  • Inferred. Some of the cost was matched from branch, commit, or conversation evidence rather than a direct link.
  • Low confidence. The evidence for the match is weak.

When Claude Code opens or pushes to a pull request during a session, Tokenize links that pull request to the session. The person needs the desktop app installed. See Install Tokenize on your computer. If nothing is linked yet, the section says "No linked pull request activity in this window."

Worth a look

The last group, Worth a look, lists pull requests that may need attention. Each card shows whether the pull request is open or merged, its AI cost, and a badge that says why it's listed, with the number behind it where there is one:

  • Reverted, or Reverted by #106 when Tokenize knows which pull request reverted it
  • Turnover and a share, such as Turnover 64%
  • Open since and a date, such as Open since Oct 2
  • Merged unreviewed
  • Rubber-stamped and the size of the change, such as Rubber-stamped · 500 lines

The group starts collapsed. Click its header to open it.

Opening a pull request from a project page shows only that one, with "Showing owner/repo#123 only." Click Show all pull requests to see the rest.

Tasks

Linear issues joined to their pull requests. It needs Linear connected; until then it says "Connect Linear to see tasks". See Connect Linear.

  • Delivery: Completed, Lead time, Cycle time, and Against due date.
  • Bugs: Bugs filed, Bug-fix share of PRs, and PRs with no issue, with Bugs filed by priority.
  • Recently completed: each task with Completed, Lead time, Due date, and AI cost. Hover a task's name to see its team and assignee. AI cost shows "—" for a task with no linked pull request.

Hotspots

The files merged pull requests changed most, with Changes, Authors, Lines, and Churn 14d.

Languages

Claude Code edits by programming language, with Accepted, Rejected, and Accept rate. Until people make edits in Claude Code, it says "No edit activity" and "Claude Code edits appear here."

Skills

Spend by skill, with Spend, Requests, and Sessions. Spend that isn't tied to a skill is shown on its own row. Admins only. Skills always covers the whole organization; with a team picked, a note says "Showing the whole organization. This view can't be filtered by team yet."

AI cost on pull requests

AI cost on Worth a look, Teams, People, Agents, and Tasks covers all the agent usage on each pull request, not only the usage in the range you picked. A cost marked ≥ is a minimum and may still rise. A cost marked ≈ may still change.

Troubleshooting

"Too many pull requests or tasks to show. Pick a shorter range or one team." The range covers too much activity. Pick a shorter range, or an admin can pick a team.

Pull requests still show under "No GitHub details". Check that the repository is listed on the GitHub card. See Connect GitHub.