Presenter playbook¶
Keep asking the same question: would you give this release more traffic? The 10-minute version tells a complete story. The 15-minute version adds depth to the same decisions.
Before the audience arrives¶
- Confirm the actual cluster baseline is healthy and traffic is flowing.
- Verify a real GPT-4o mini investigation and all required tools.
- Check the current operator's controller labels and log access.
- Rehearse clearing the fault and draining the 60-second observation window.
- Prepare the explicitly labeled local rehearsal as a fallback.
- Open the dashboard and Argo CD with
./demo.sh present --observability. Complete OpenShift SSO before going on stage; selectcanary-checkoutand leave its resource tree open in the second tab. - Select Core or Extended in the SPA. Present the flow at your own pace; there is no presentation timer.
The 10-minute core¶
| Time | Scene | Audience question |
|---|---|---|
| 0:00–1:00 | SPA baseline → Argo CD Git revision/resource tree → start v2 | What evidence would justify more traffic? |
| 1:00–3:30 | Shared slowdown, HOLD, recovery, GO → Argo CD successful analysis/promotion | Did this release cause the latency increase? |
| 3:30–7:30 | v3 retry regression, NO-GO → Argo CD failed analysis/abort | Can successful requests conceal extra dependency work? |
| 7:30–9:00 | Evidence and policy record | Why should the platform trust this recommendation? |
| 9:00–10:00 | Git change and shared template | What does an application team need to adopt this? |
Model calls, controller reconciliation, and observation windows consume real time. Measure these on your cluster. If the timing cannot fit, shorten optional narration or enter at the prepared retry scenario; do not weaken the evidence or bypass the gate.
Add five minutes¶
Two minutes: follow the investigation¶
Expand the agent workflow diagram. Use about 30 seconds for the Prometheus comparisons, then 90 seconds to open one request observation, follow its trace ID in Jaeger, and match it to the retry log record. Show the relevant @Agent and @ToolBox definitions.
Three minutes: govern autonomy¶
Apply Production policy to a healthy candidate. Show a GO recommendation held pending approval. Demonstrate that missing evidence still blocks promotion, even when someone wants the rollout to proceed.
Tell each audience what changed¶
Platform engineers: the organization provides a shared investigation service, central policy, scoped access, and a reusable AnalysisTemplate.
AI engineers: the system composes bounded specialists, preserves evidence identity and memory isolation, validates tool use, and records uncertainty.
Developers: a small Git release change produces a decision they can inspect, with an explanation of the observed application behavior.
Rehearse failure as deliberately as success¶
Verify healthy candidates, shared latency, canary retries, missing traces, sparse traffic, expired evidence, repeated releases, wrong-stage decisions, model outages, and unavailable controller APIs.
Measure recommendation quality, evidence-reference validity, investigation latency, and tool count separately. The current records do not expose a complete token-cost accounting system, so do not present an invented ROI or confidence score.
Keep the scope honest¶
This is a bounded decision loop in one cluster. It demonstrates a platform contract; it does not establish multi-cluster scalability, universal causal diagnosis, durable audit retention, or production reliability guarantees.
Shared tools and policies explain how teams can adopt the approach. Add full trace-tree queries, an external log adapter, a versioned evidence archive, and policy distribution when those additions strengthen the decision or developer experience.
Argo CD UI checkpoints¶
Keep Argo CD open in a second tab. In the core demo, spend about 20 seconds identifying the Git revision and Rollout at baseline, then 20–30 seconds on the AnalysisRun and stable ReplicaSet after NO-GO. While the recovered observation window is warming, use Argo CD to show that a pending/inconclusive analysis has not authorized promotion. After GO, show a Successful AnalysisRun and progression to the next step. After NO-GO, show the Failed AnalysisRun and aborted/degraded Rollout while the stable ReplicaSet remains available. Synced means desired configuration matches Git; it does not mean the canary passed health analysis.
In the 15-minute version, add one minute to inspect the reasoning-gate AnalysisTemplate, its web-provider URL, and the AnalysisRun result. Take that minute from the extended investigation walkthrough; the production approval and missing-evidence scene remains three minutes. This makes the separation visible: Argo CD reconciles Git, Argo Rollouts executes progressive delivery, agents recommend, and Decision Policy authorizes the gate.