Trust the output, sheet 03 of 13
Metrics for an org where agents author a growing share of the code, and an honest counter to a top-down mandate.
Pair cost per merged PR with incidents per merged PR. Nobody sells a pushback metric yet, and your GitHub event stream already has the data.
← → move to the next sheet over, Esc zooms out.
Once agents write a real share of the code, the old dashboards keep counting and stop meaning much. More pull requests, more lines, more tokens. None of it says whether the work got better, or how often a person caught a bad change and stopped it.
It's like judging a kitchen by how many plates left the pass. Useful, until you notice nobody's counting how many came back.
The honest instruments come in pairs: what it cost and what it broke, what shipped and what got rejected. Most of the data already sits in your Git history. Nobody packages the uncomfortable half.
Every link opened on 11 Nov 2026. The numbers match the squares on the map.
Reference, DORA
The long-running research program on software delivery performance, with its 2025 report on what AI-assisted development changes.
The catch. Survey research. It tells you what correlates across thousands of teams, not what happened in yours.
Reference, METR
A randomized controlled trial: experienced open-source developers took 19% longer to finish issues when they could use AI tools.
The catch. A snapshot of early-2025 tools. METR's 2026 follow-up found returning developers about 18% faster, with a confidence interval from 38% faster to 9% slower, and METR calls the new data an unreliable signal. Read it for the gap between perception and reality.
Framework, DX
Research-based metrics for measuring the impact of AI code assistants and agents on an engineering org.
The catch. A framework from a vendor that sells the measuring. Use the questions even if you don't buy the platform.
Reference, GitClear
A multi-year look at changed lines: refactoring down, copy-paste up, and short-term churn rising as AI assistance spread.
The catch. GitClear sells the analysis. The trend is useful, and your own repo is the better evidence.
Product
Traces coding-agent usage and token spend to shipped features and incidents, so cost and outcome sit in one place.
The catch. The figures in its reports are Faros's own. Check them against your incident log.
Product
Team-level engineering analytics that flag early signs of AI slop, such as falling code quality or human review missing issues in AI changes.
The catch. Leading indicators are prompts for a conversation, not grades for a person.