Know what your engineering org actually ships.
Weave looks at every PR and answers one question: how long would this take an expert engineer to complete?
Output trend across the last 13 weeks, benchmarked against peer engineering teams.
What output measures
One standardized unit of work, applied to every engineer, every repo, every language.
One standardized unit of work
Consistent and comparable across individuals, teams, languages, and organizations.
Meaning over volume
Unlike lines of code or PR count, Weave analyzes the impact of every change, not just how much was written.
Apples to apples, every time
The same model runs across every engineer, every repo, every language. No exceptions.
Compare teams
Visualize output across squads. Identify high-performing teams and surface where technical debt or process friction is slowing delivery.
Output per engineer · 30 days
Individual drill-down
View output per engineer within a team context. Spot coaching opportunities and recognize your most impactful contributors.
Platform · by engineer
| Engineer | Output | Quality | Reverts |
|---|---|---|---|
| M. Chen | 142 | 94 | 1 |
| R. Mehta | 131 | 92 | 0 |
| J. Park | 118 | 89 | 2 |
| S. Tombs | 96 | 91 | 1 |
How Silk 1 estimates expert effort
Weave's Silk 1 model combines an LLM, static analysis and a custom model trained on PRs hand-labeled by the engineers who wrote them. It follows dependencies across files, so a short change to a shared interface can score higher than a long, mechanical diff. A score of 3.0 represents three hours of estimated expert effort. The figure is a modeled measure of the work. It does not record the time an engineer spent or saved.
Every score carries its reasoning. If the author knows a change was harder or easier than the model assumed, a team score override corrects it and feeds organization-aware calibration. The Silk 1 release reports a 3.3x reduction in average prediction error against the previous model on Weave's evaluation set. Teams can compare both versions side by side, with historical backfill.
What quality tracks
Three signals that connect production incidents back to the code that caused them.
Trace bugs to the PR that introduced them
Every incident links back to the commits that shipped before it. Stop reconstructing the timeline by hand.
Incident → cause
- Incidentsev-2 · 482
- Deploy2.14.0
- Pull request#1204
Track revert rates
See which engineers and repos have the highest revert rates. Catch quality issues before they compound.
Revert rate by repo · 90 days
Identify volatile codebases early
Spot areas of your codebase with high churn and rising bug rates before they become costly to maintain.
Churn by week · payments
What reviews measure
Metrics that show you whether your review process is a quality gate or a bottleneck.
Choosing an automated reviewer? Compare AI code review tools and evaluation methods.
Review quality
Score eligible reviews on depth of understanding, communication, practicality and impact: whether feedback led to meaningful code changes. This evaluates the review, not the entire PR, and cannot detect errors the reviewer omitted.
Average review cycles before merge
Track how many rounds of review each PR goes through. Identify bottlenecks and reduce the cycles that slow delivery.
Review volume by team
Break down review volume by team member and week. Balance workload and keep reviews from piling up on your most senior engineers.
Rounds to merge · all repos
Reviews per reviewer · weekly
What benchmarks cover
Benchmarks calculated from real engineering data across thousands of organizations — not a published report.
Deploy MTTR · benchmark tiers
- EliteUnder 1 hour
- HighUnder 1 day
- MediumUnder 1 week
- LowOver 1 week
CI/CD metrics
Core deployment health
Deploy MTTR
Mean time to recovery from a failed production deployment. Benchmarked from elite (under 1 hour) to low (over 1 week).
Deploy frequency
How often your team deploys to production.
Deploy success rate
Percentage of production deployments that succeed.
PR deploy lead time
Time from code merge to production deployment.
Industry benchmarks
Segmented by org size: 1–5 up to 201+
Code output per engineer
Compare normalized output against real data from thousands of engineering orgs.
PRs per engineer
Fair comparisons regardless of your team’s scale.
Code review quality & turnaround
Industry-standard review depth and cycle time benchmarks.
Code LOC turnover
How much of your codebase is being rewritten relative to industry peers.
What AI impact tracks
Two signals that go beyond prompt counts, connecting AI activity to what actually ships.
Know how much of your code is written by AI
Track AI-generated code as a percentage of total output. Benchmark against industry peers so you know where your organization stands.
Authorship of merged code · 30 days
See the impact of AI on your output
Compare AI-assisted code against actual engineering output across bugs, features, and maintenance. Volume alone doesn’t tell the story. Weave connects AI usage to what actually ships.
AI-assisted share by work type
AI contribution and cost with visible evidence
Weave attributes AI-assisted code using the strongest evidence available for each change, in this order:
- direct tool API data
- line-level git-ai notes
- Git co-author metadata
- activity inference
Engineers still receive output credit for work they drove with an agent. AI attribution describes how the code was written, and output credit describes who did the work.
Token Intelligence reports AI usage and cost per person, tool and model where the connected source supports that detail. Each cost is labeled as billed, list-price estimate, plugin-observed or activity only, so finance can see which figures match an invoice.
What the score tracks
Three signals behind the score.
Weave syncs usage metadata from connected IDEs and CLI tools. It attributes code output to either AI or human authorship using advanced heuristics, then correlates AI usage with engineering productivity and cost data to surface the full picture.
Foreground vs. Background AI
Categorize your tools into foreground AI (interactive tools like Cursor and Claude Code) and background AI (automated agents). See how each category impacts your codebase differently.
Impact tracking
Visualize how specific tools influence delivery speed. Correlate adoption of Cursor, Windsurf, Claude Code, and others with actual productivity changes across your team.
Financial health
Monitor subscription spend per user alongside output data. Make sure your investment in AI tooling aligns with the productivity gains you’re actually seeing.
Investigate a review queue after an agent rollout
This example is illustrative and does not describe a customer.
After a team adopts coding agents, PR count rises and review turnaround lengthens. Sorting the period's PRs by estimated effort in Weave shows that most new PRs are low-effort generated updates. A few high-effort cross-file changes waited longest and received shallow reviews.
The team responds in three steps:
- Assign an experienced reviewer to cross-file changes.
- Write the team's standard for generated updates as a Weave Check that runs in CI.
- Check review depth and turnaround again the following month.
Customer results and investment reporting
Reducto reports a 19% increase in output per engineer measured in Weave.
Bytes AI reported about 30% more engineering output six weeks after opening every engineer's Weave profile to the whole team.
Intellistack connected GitHub, Cursor and Claude Code to build a historical view of AI adoption for board reporting. At PostHog, more than 80 engineers inspect their own Weave data and adjust how they work.
Connected data and access
Which engineering data can Weave connect?
- Source control: diffs, commits, reviews, comments and merges from GitHub, GitLab, Bitbucket, Azure DevOps and Gerrit.
- Planning and delivery: issues and deployment events from connected trackers and pipelines.
- AI tools: usage and cost from tools including Claude Code, Cursor, Copilot, Codex, Devin and Windsurf. Coverage varies by tool.
- Attribution and sessions: git-ai line notes, co-author metadata and, when the Weave plugin is installed, coding-agent session data.
Can engineers and agents query the underlying records?
Wooly, Weave's research agent, answers natural-language questions across output, reviews, deployments, AI activity and cost. Each answer links to the PRs and records behind it. The read-only MCP server brings the same data into Claude, Cursor and other compatible clients. Permissions can limit members to their own metrics while keeping organization aggregates visible.
The engineering intelligence platform for the AI era.
Output, quality, review health, and delivery benchmarks in one place — measured consistently across every team you run.

