Show, don't promise.

Stand up the platform on Workers, Durable Objects, Workflows, Queues, AI Gateway, Vectorize, D1, R2 — used the way Cloudflare intended.

• Monorepo layout (apps/, packages/, infra/)

  agent-router, agent-basal, config-api, fleet-api, workflow-control, workflow-runner, shared, agents-sdk-glue.

• Each Worker scaffolded with proper bindings

  workers_dev=false, observability on, no hand-rolled JSON-RPC. Reviewed by DevOps before any deploy.

• Existing /api/chat path migrated through agent-router

  Behind a feature flag defaulting OFF — production behavior unchanged until flipped.

• D1 schema + Terraform plan only

  No migrations applied, no terraform apply, until reviewed in the morning.

Every page exists with full Shannon-reviewed copy and mock data. The vision and governance are reviewable by leadership before backend wiring lands.

• /vision

  Highest-value-work principle, fear-and-answer pattern, what's changing / what isn't.

• /governance

  Roles, permissions, kill-switch, audit, cost caps, PHI scope, change control, honest gaps.

• /agents, /agents/me, /chat, /workflows, /fleet, /fleet/briefing, /onboarding, /roadmap

  All routes live with mock data; structure validated end-to-end.

• /styleguide updated with new components

  FearAndAnswer, RoleBadge, AuditRow, CostMeter, FleetTile, WorkflowCard, BriefingPlayer, AgentConfigCard, etc.

Real per-user agent configuration, real memory, real audit log — written through the config-api Worker and the agent's Durable Object.

• D1 schema applied

  users, user_agent_config, audit_log, chat_threads, chat_messages_index, kill_switches, cost_caps, workflows, workflow_versions, workflow_runs, briefing_episodes.

• agent-basal Worker + DO live

  Cloudflare Agents SDK; per-user thread + memory pointer.

• Vectorize index for memory

  Save/recall round-trip; 'show what this agent remembers about me' surface real.

• /agents/me writes through config-api

  Toggle, custom instructions, memory delete — all audit-logged.

Cloudflare Workflows execute the seeded list. Fleet aggregates from real audit data. AI Gateway in front of every model call. Kill-switches enforced server-side.

• Audit ingestion webhook

  ai-proxy posts to /api/audit/ingest; D1 + Analytics Engine populated.

• /fleet aggregates

  team-only by default; ai-admin sees full; per-employee leaderboards never.

• Kill-switch round-trip

  KV flag, defense-in-depth check at router and at agent Worker.

• Cloudflare Workflows runtime

  Seeded workflows: morning briefing, support triage, daily fleet briefing, weekly company briefing.

• /chat persisted threads + R2 attachments

  Per-user DO threads + signed-URL upload to R2.

• AI Gateway in front of Anthropic

  Caching, observability, per-agent budget enforcement.

Workflows are designable in the UI; the daily briefing has a real podcast feed; first-visit onboarding is wired.

• Workflow designer

  Edit steps, save versions, A/B promote a candidate version when it wins.

• Briefing TTS pipeline + RSS

  Audio MP3s in R2; subscribable podcast feed at /fleet/briefing/feed.xml.

• First-visit onboarding redirect

  Role-based path triggered until completed; never fires again.

• PHI-authorized gate enforced

  All agents tagged piiScope:phi require the phi-authorized group.

• Champions opt-in directory

  Per-agent 'ask me about this' surface on each agent page.

Important, real, scoped — and parked until we're ready to ship them with counsel review rather than as filler.

• CCPA employee data export endpoint

  Self-serve 'export everything stored about me.' Hooks are in the schema; export endpoint awaits counsel pass.

• AB 2930 automated-decision flag

  Mandatory human-in-the-loop documentation for any agent that affects employment decisions. We don't run any such agents today.

• Surveillance policy text

  Explicit, written-down policy on what the Fleet view is and isn't. Drafted; needs counsel review.