43,280 AI events recorded on our own engineering fleet in the last 30 days.
One shared policy model for the coding tools in use, so developers know what is allowed and platform teams see what needs attention. Then measure what changes: delivery, quality and standards adoption, computed from the same activity.
TRACELET.AIEngineeringIllustrative exampleEnrolment & activity
Enrolled machines and the AI activity recorded on themNeeds your attention
Fleet, policy and tool items ranked by urgencyCheck the Tracelet agent on each machine before the next policy update.
4 supported actions were blocked this period. Review the decision records.
Detected on the Data platform team. Approve it or add it to the policy.
Teams by coverage
Enrolled machines and the next step for each teamIllustrative example: tools, policies and fleet health for the machines your developers work on.
Check tool coverage before planning the first stage.
Enrol machines and confirm which coding tools, MCP servers and actions are covered for each team.
Set the first policies for file, shell and MCP actions. Developers see the policy and the reason where the action happens. Publish the rules that work as standards and push them to every repository they apply to.
Explore software delivery policiesOne team and a small set of policies first. Decision records and fleet health decide where to expand next, and delivery and quality metrics show whether each stage is paying off.
The standards that turn useful AI practice into a supported way of working.
Illustrative exampleBefore an agent makes a change, it reviews the relevant files, recent commits and open work.
Evidence and review are complete.
Every AI-authored change carries an intelligible reason and a linked work item.
Two teams are validating this pattern.
AI tools can investigate production, but sensitive credentials never leave the approved path.
3 requests stopped before credential access.
Delivery velocity, pull request activity, issue throughput, and CI pipeline and release health, computed from ingested GitHub history. AI-tool coding activity is plotted on the same time axis, so the question of whether AI use tracks with delivery is answered from one report.
Merges to the default branch counted as deploys, 12-week average.
+3 per week vs prior 12 weeksAverage hours from pull request opened to merged.
−6 h vs prior 12 weeks13 of 168 counted deploys linked to a revert or an incident-labelled issue.
Structural proxy, not ground truthAverage hours from incident-labelled issue opened to resolved.
Structural proxy, not ground truthChange failure rate and MTTR are structural proxies derived from linked reverts and incident-labelled issues in the ingested history. They are not ground truth for production incidents.
Performance bands are derived per developer from keep rate, session frequency and feature depth, then reported only as counts per band and per team. Individual rankings are kept out of the report by design.
TRACELET.AIEngineering / DevelopersIllustrative exampleFleet distribution
Share of developers in each performance bandKeep rate above 75%, eight or more sessions a week, deep feature use.
Steady sessions with keep rate around the fleet median.
Few sessions or a low keep rate in this period.
Movement vs prior 30 days: +4 into High · −2 out of Low
What feeds the banding
Three signals, each computed per developer and reported as a fleet medianShare of AI-suggested code still present at merge.
Active AI coding sessions per developer, per week.
Distinct capabilities used in a session: edits, tests, MCP, review.
By team
Aggregate bands per team, with counts instead of namesAggregate by design. Bands are computed per developer but never listed by name in reports, and any group smaller than five developers is suppressed.
The catalog holds the organisation’s recommended skills, rules, tools and MCP servers. A rollout fans one standard, or a whole pack, out across repositories as tracked pull requests.
4 of 5 pull requests opened · 1 merged · 1 pending
acme/payments-apiSourceacme/checkout-webMergedacme/ledger-servicePR #88 openedacme/notificationsPR #231 openedacme/billing-workerPR #57 openedacme/identity-gatewayPendingPlatform baseline · 6 standards · clone the whole bundle onto acme/new-service
Read down a column for a team, across a row for a dimension. Each cell is scored against the row’s threshold, so the gaps that need a standard or a policy stand out without a separate report per team.
TRACELET.AIEngineering / QualityIllustrative exampleRegistered projects the fleet’s AI activity rolls up to, and where that activity is happening by working repository. Most projects are auto-discovered from detected AI activity on a repository rather than registered by hand, and near-duplicates can be merged.
Two auto-discovered projects point at repositories with the same name stem. They are selected below; merge them to keep one project and its history.
github.com/acme/payments-apiHealthy$1,24012318github.com/acme/checkout-webNeeds attention$86082043 repositories · registered by handHealthy$2,0506142github.com/acme/ledger-serviceHealthy$410596github.com/acme/notificationsNeeds attention$120341github.com/acme/notifications-svcNo activity$3516github.com/acme/identity-gatewayNo activity$000Showing 7 of 38 projects · auto-discovered projects are created from detected AI activity on a repository