SDLC · Stage 06 of 07
Operate and observe
You should be able to follow any change from start to finish in one place, and trust the numbers you see.
Principles: Event-driven by default, Interfaces people trust
When something goes wrong at 2 a.m., the platform should help you understand what happened, not make you piece it together from five different tools.
Every pipeline step emits an OpenTelemetry span, and the spans are stitched into one Tempo trace per end-to-end flow, even though each stage is a separate PipelineRun. Grafana shows the live and historical pipeline list, a per-stage drill-down and DORA metrics computed from the CDEvents stream. MTTR is shown as experimental, because a rollback done outside the pipeline is a blind spot that would otherwise make it quietly wrong.
Run history is archived with Tekton Results; stalled pipelines are detected and alerted on; old runs are pruned. Tower brings it together for the people running services: releases per environment, rollout topology, SLO burn rates and fleet-wide views.
When a rollout degrades, the control plane itself asks for a diagnosis and comes back with a fix PR.
Where Hangar does this
And for AI agents
A flight recorder for agents, and alerts that arrive with a tested hypothesis
Agent observability reuses the same telemetry and adds one tamper-evident, hash-chained audit trail. Triage moves from a bespoke integration to a bounded agent behind Clearance, and remediation earns autonomy one alert class at a time.