Skip to content

Plan to 26 October

Built on the findings in the Premise review. Roughly eight weeks from 2026-09-02.

For the two of us

This page is the what. The Development plan is the how: the hypotheses we need to prove, the eval harness that proves them, how the work splits between us, and the decision gates.

Decisions to make first

These block everything else. Take them in week 1.

# Decision Recommendation Cost of delay
B1 Cluster: which environments do we support? All three — Codespaces (default), local k3d, BYO — on one namespace-scoped contract Makefile and all docs need rewriting
B2 Agent runtime: custom code or Claude Code + MCP? Claude Code + K8s MCP We build a framework we can't test in time
B3 Model access for participants Workshop keys with a spend cap The room stalls at minute 30
B4 Lab 3: hands-on or demo? Demo We overrun and lose the wrap-up
B5 Website: one or two? Astro at root, MkDocs at /docs Two designs, two domains, confused participants
B6 Do we run a quest board? Yes, by forking the MCP workshop's quest-server — built last, cuttable We build a scoreboard from scratch, or bore the fast third of the room

Revised schedule

Adjusted for buffer, and for Lab 3 becoming a demo. Still 120 minutes, but with slack.

Time Duration Block Change from proposal
00:00 12 min Intro: What is Agentic Cloud Ops? −3, tightened
00:12 8 min Bootstrap + verification −7, assumes pre-pull at home
00:20 5 min Buffer / help stragglers New
00:25 30 min Lab 1 — Your first agent unchanged
00:55 35 min Lab 2 — Security and hardening unchanged, protected
01:30 10 min Lab 3 — Chaos, as a demo −5, format changed
01:40 5 min Buffer New
01:45 15 min Q&A and wrap-up +5

Why bootstrap shrinks rather than grows

Because the work moves home. Block 2 should be "run one command, see two green nodes" — not "download Kubernetes". See P1.


Phases

Phase 1 · Decisions and skeleton (weeks 1–2, September)

  • Take B1–B6
  • Build bootstrap/base + make verify — the environment contract — and get it green on arm64 and amd64
  • devcontainer.json + prebuilt image, so the Codespaces path is proven early
  • make preflight that also pre-pulls images and fails clearly
  • Wire the MCP server to the cluster; verify the agent can see nodes
  • Ask the organisers about the P9 items

Phase 2 · Labs 1 and 2 (weeks 3–5, Sept–Oct)

  • manifests/insecure/ — the deliberately bad deployments
  • System prompts: deploy.md, harden.md, incident.md
  • Lab 1: the planted ImagePullBackOff, and verify the agent actually finds it
  • Lab 2: make health as a deterministic, external health check
  • Write down expected agent behaviour per lab, for the facilitators
  • Checkpoint branches: checkpoint/lab1, checkpoint/lab2

Phase 3 · Lab 3, evals and content (week 6, October)

  • manifests/chaos/ — pick one deterministic scenario for the demo
  • Re-scope dns-mess and node-pressure to namespace level (capability matrix)
  • Fork the quest server; six quests sealed under a private pepper
  • Evals: run Labs 1 and 2 ten times each, measure success rate. Below 80% → simplify
  • Slides for blocks 1 and 6 — including CVE-2026-46519 (P5)
  • Landing page in Astro (spec)

Phase 4 · Dress rehearsal (week 7, ~19 October)

  • Full timed run-through, both speakers, from a cold machine
  • Test on a machine that isn't the development machine — ideally a work laptop behind a proxy
  • Freeze the prompts. No changes after this point
  • Write the facilitator runbook: common failures and what to say

Phase 5 · Final week (week 8)

  • Send participant instructions with pre-pull at the top, at least 5 days ahead
  • Generate workshop API keys with a spend cap; print them on cards
  • Test on the conference network the evening before, if possible
  • Upload slides to the CFP portal (still outstanding)

Still missing from the repo

bootstrap/           base kustomization + the three env scripts
evals/               the harness, scenarios, checkers and results
quests/              forked quest-server, quests.yaml, answers.yaml
agents/prompts/      deploy.md, harden.md, incident.md
agents/tools/        MCP config, RBAC for the agent's service account
manifests/insecure/  root containers, no limits, no NetworkPolicy
manifests/chaos/     at least one deterministic scenario
Makefile             agent, agent-check, lab1-break, health, chaos (all TODO)
web/                 the Astro landing page

Risk register

Risk Likelihood Impact Mitigation
Conference network can't carry the room High Critical Pre-pull at home; images as a hard prerequisite
Participants without model access High Critical Workshop keys on cards
The agent behaves differently across the room Certain Medium Checkpoints, evals, make it the curriculum
We run out of time before block 6 Medium High Two buffer blocks; Lab 3 as a demo
Lab 2 breaks the app for everyone Medium Medium make health + rollout undo in the prompt
Work laptop blocks Docker or the API Medium Medium Codespaces as the default path
Quest board eats time the labs needed Medium High Built last, after Gate B; cuttable to six quests
BYO participant points the agent at something real Low Severe byo.sh guard + namespace-scoped RBAC (Environments §5)