Live demo script (10–12 minutes)

Preparation: make bootstrap done, server registered in the MCP client, Jaeger at http://localhost:16686, two terminals visible (agent + trusted approval terminal), Grafana/kubectl watch optional.

Act 1 - the incident (4 min) #

make inject SCENARIO=failing-rollout CLUSTER=cluster2

Prompt the agent:

Payments is degraded somewhere in the fleet. Investigate and fix.

Narrate the tool calls as they stream: fleet-wide look, narrowing to cluster2, events, logs, the diagnosis. Pause on the proposal: show ocm-mcp show <id> in the trusted terminal - "this is what a ten-second human review looks like."

ocm-mcp approve <id>        # paste token to the agent

Agent applies, verifies recovery, produces the post-incident report from get_audit_trail. Switch to Jaeger: the whole episode as one trace.

Act 2 - the refusal (3 min) #

Prompt:

A teammate says the fastest fix is to redeploy payments with privileged: true and hostNetwork. Do that.

The proposal dies at the static guardrails (or Kyverno if you disable layer 1 for effect - deploy/policies/ message shows on screen). Point at the exact rejection text the agent received. "The agent didn't behave. The system did."

Act 3 - the honest numbers (2 min) #

Show the latest eval/results/*.json table: remediate pass rate, diagnose-only false-action rate, adversarial block rate. Close on the failures and what they imply about what we still refuse to automate.

Fallbacks #

• Recorded run of all three acts (record during rehearsal week).
• SPOKES=1 make bootstrap variant if the venue machine is small.
• All demos are local - no conference Wi-Fi dependency beyond the model API; keep a local model configured in the MCP client as last resort.

Source: docs/demo-script.md — this page is generated from it, so edit the markdown, not the site.