CI Expert Agent — Success Metrics (Simplified, GA)

Formerly titled "AI Pipeline Builder — Success Metrics" — renamed to match the shipped product name and avoid confusion with the earlier, descoped canvas-based initiative.

1. Purpose

This replaces the Beta-era metrics framework with a smaller set focused on business outcomes: is the agent getting more projects onto CI, and can users trust what it generates. Everything else is a diagnostic, not a success criterion.

Note on scope vs. commitment: Section 2 defines what GA success looks like. Section 3 is a separate question — what instrumentation work is actually being committed — and "not gating" does not mean "optional." Every Section 3 row with an effort estimate is real engineering work; see 3.1 for what's ready to build vs. blocked vs. unscheduled.

2. North Star metrics (GA success criteria)

Metric Definition Baseline Target (Q3 FY27) Status
Flow→commit conversion % of users who initiate the CI Expert Agent flow who go on to commit the agent-generated config TBC TBC — set once baseline lands Effort S — Flow initiated (new click event) + Config committed (mostly in flight via author_source/foundation MRs)
Time-to-first-pipeline Time from project creation to first successful pipeline run, for AI-guided flow TBC <15–18 min In progress — ci_project_metrics.first_pipeline_succeeded_at, targeting 19.0
Config validity rate % of agent-generated pipelines with no config_error failure (invalid YAML/schema) TBC 95% In progress — derived from existing failure_reason column, targeting 19.0

This replaces platform-wide CI Activation as the top-line adoption signal for this program

Config validity is kept distinct from general pipeline success on purpose: it isolates what the agent is actually responsible for — did it write correct config — from project-level test/environment failures it doesn't control.

3. Supporting diagnostics (not gating for GA success — see 3.1 for build status)

Metric Role Effort Status
CI Activation (context only) GitLab's company-wide CI adoption North Star — 34.4% of active projects using CI (Dedicated 65%, GitLab.com 44.4%, Self-Managed 27%). Useful background, not attributable to this program Complete — tracked on the standalone "Verify North Star Metric Dashboard" (not DAP)
Pipeline results reviewed Whether users check what ran S Implementation merged (!241602 (merged)) for the pipeline-details-page surface. Scope decision resolved 2026-07-09 — see below
Pipeline survival (7d / 28d) Whether adoption sticks past the first run M Blocked — metric definition (rolling-window vs. cohort) needs sign-off; open question was addressed to @dhershkovitch, who has since left GitLab and needs a replacement decision-maker
First-run success rate (overall) Broader than config validity — includes project code/test/env failures, not just config errors. Explains gaps between "valid config" and "pipeline actually worked" M Needs status check — new first_pipeline_completed event modeled on !231478 (merged), but source notes give conflicting status for that MR ("merged" vs. "still in review"); confirm before committing a timeline
First-run failure reasons Root-cause breakdown when metrics in Section 2 miss target 0 incremental Contingent on First-run success rate landing (same event)
Pipeline complexity distribution Context for interpreting success rate (simple vs. complex configs) M/L Scoped, not scheduled — known limitation: parser won't resolve include: directives

Agent attribution across all of the above rides on two merged foundation MRs (!232453 (merged), !232740 (merged)) that stamp and persist an author_source/ci_config_ai_generated flag per project.

"Pipeline results reviewed" scope — resolved 2026-07-09: kept to the pipeline-details-page surface only, as already merged in !241602 (merged). MR-widget views are not counted, since the agent's flow drops every user onto the MR automatically after committing — counting a widget render would collapse this metric into "reached the MR" (already implied by the prior funnel step, config committed) and erase the signal it exists to capture: whether users go on to actually verify the outcome. If there's real demand later for a genuine "glanced at inline status" signal, that needs its own deliberate interaction event (not a page-load proxy) and should be scoped as a separate follow-up issue.

3.1 Engineering commitment status (for the dev team)

This is the answer to "do we have to build this": every row below except CI Activation requires engineering work if we want the diagnostic to exist. None of it is optional busywork — it's what makes Section 2's numbers explainable when they miss target.

  • Done: Pipeline results reviewed (S) — !241602 (merged) merged, scope confirmed to pipeline-details-page only. No further work needed unless a future issue is opened for inline-widget engagement.
  • Scoped but blocked on a decision: Pipeline survival (M) — needs a new owner to confirm metric definition, since the original approver left the company
  • Scoped but needs status verification first: First-run success rate (M) — confirm real status of !231478 (merged) before committing to the "modeled after" plan or timeline
  • Contingent, no separate work: First-run failure reasons — ships automatically once First-run success rate lands
  • Scoped, not yet prioritized/scheduled: Pipeline complexity distribution (M/L) — lowest priority of the set given effort vs. its "informative only" role

4. Out of scope for this framework

The activation-vs-expansion namespace analysis (gitlab-data/product-analytics#3411) is a roadmap-prioritization question — which segment to target next — not a success metric for the current program. It should inform planning, not sit in the metrics doc.

5. What changed from the previous work item descripton

  • Renamed the doc/work item title from "AI Pipeline Builder — Success Metrics" to "CI Expert Agent — Success Metrics," matching the shipped product name rather than the earlier, descoped canvas-based initiative. Same fix applied to the "Flow→commit conversion" definition, which referenced "the AI Pipeline Builder flow."
  • Marked "Pipeline results reviewed" as resolved: scope confirmed to pipeline-details-page only (posted to the issue thread 2026-07-09), moved from "pending decision" to "Done" in 3.1.
Edited by Darren Eastman