The nine deep dives we deferred.

The Agent-Native OS is an 8-layer stack. The diagram in architecture/AI-BUILD-LAB-OS-STACK.md (HQ repo) is the canonical reference. Layers (top-down): Presentation, Routing, Knowledge, Trust, Agent Fleet, Tooling, Memory, Substrate.

4D students touched layers 1, 5, and 8. 8D students reached into 2, 3, and 6. The remaining layers ship as you grow your setup.

A Cairn is a per-source three-tier knowledge package: L1 waypoints (the route map), L2 cards (Chain-of-Density summaries), L3 raw evidence. The Federated Librarian agent fans out across all your active Cairns to answer queries.

Existing student-repo guides at guides/cairns/00-what-is-cairns.md through 06-agent-skill-templates.md walk through the setup. Use them in order.

The 4D path uses 1Password with the op:// URI pattern: store the key in 1Password, reference it via op://Vault/Item/field, run with op run -- <your-command>. The key never lands in .env or shell history.

Without paid 1Password (CLI is paid-tier only), fall back to a gitignored .env. Confirm .env is in .gitignore before any commit.

For the 8D path, you layer in: Infisical (multi-vault for team secrets), Proton Pass (cross-platform alternative to 1Password), and a key hierarchy (root keys vs scoped per-app keys).

Three ways to reach Claude:

• Claude Code (CLI): terminal app, lives in your repo, edits files directly. Day-to-day driver for cohort 1.
• MCP (Model Context Protocol): tool-server protocol. Claude calls registered tools; you write the tool spec. Use when Claude needs to talk to an external service.
• SDK: programmatic API. You write code that calls Claude inside your own app. Use when Claude is a feature inside something you're building.

Decision: file editing → CLI. External service → MCP. Embedded in your product → SDK.

Multiple Google accounts. Proton + Outlook + Gmail. Different mailboxes, different OAuth scopes, different tokens. The setup is fiddly; treat each account as its own connector with its own scopes.

Vercel hosts. Supabase stores. The morning-brief 4D scaffold uses Vercel for hosting + serverless functions and Supabase for any persistence you need. The 8D extension menu shows the database wiring.

For your own projects: same pattern. npx create-next-app → vercel deploy → add Supabase as a connector.

References: blueprints/morning-brief/QUICKSTART-4D.md for the deploy flow, EXTENSIONS-8D.md for database wiring.

Three patterns:

• Patchwork: one agent, many small tasks, no sub-agents.
• Worker Swarm: lead agent writes prompts, worker agents execute in parallel, lead synthesizes.
• Research Swarm: scan-driven, manifest of blocked-by waves, agents discover and feed each other.

Cohort 1 didn't ship multi-agent in the workshop. Tyler's Etsy demo (~4:07 main room) was the closest preview.

Reference: student-repo guides/09-autonomous-workflows.md, Sara curriculum 12-progressive-autonomy.md.

Two paths people end up considering. CC OS (Claude Code OS) is the default path the workshop teaches: Claude Code as the editor, agents in .claude/, MCPs for external services. OpenClaw is the open-source alternative: more configurable, more setup, less polish.

4D is "you can build something useful with Claude in a day." Morning brief deployed. Most people stop here for a while; that's fine.

8D is "your build extends itself." Scheduled cron, Slack relay, multi-source aggregation. You start to see Claude as infrastructure rather than a tool.

12D is the next horizon beyond 8D: agents that operate semi-autonomously across a fleet of repos and surfaces. The gates between 8D and 12D are still being mapped.

References: blueprints/morning-brief/QUICKSTART-4D.md, EXTENSIONS-8D.md, curriculum/CURRICULUM-OVERVIEW.md.