AI Superusers Club  ·  April 2026

State of My System

Dee (Dieter Werwath) Q3 2025 → Apr 2026 Architecture + Journey
01

Journey Timeline

01
Q3–Q4 2025 The Tinkerer
ChromeOS. Google Apps Script. Manual everything.
  • Built GAS automations for recurring tasks — email parsing, sheet updates, notifications
  • Correct call for the moment: fast to start, zero infrastructure overhead
  • Hit the ceiling quickly: no agent integration, brittle scripts, no path to scale
What it cost: Time. And the false confidence of "automation" that wasn't really automation.
02
Feb 2026 The Catalyst
OpenClaw arrives. First real needle-mover.
  • First WA message from OC: Feb 3, 2026
  • Within weeks: compelling enough to end 10 years on ChromeOS and return to Mac
  • First taste of agents operating across surfaces without manual triggering
What broke: Nothing yet. The honeymoon phase.
03
Mar 2026 Building Fast
Mac. Multiple projects. Infrastructure everywhere.
  • True Torque built: Python webhook server routing WA + Telegram to agents
  • Cloudflare stack expanded to 7 active projects across Workers / Pages / D1
  • Constitution drafted: governance, hard constraints, proof discipline
  • Agent hierarchy built out: orchestrators, secretaries, overseers, domain agents
What broke: The agent hierarchy. Too complex, too early. Dismantled within weeks.
04
Mar–Apr 2026 Forced Evolution
Anthropic restricts OpenClaw from subscription plans.
  • Lost the primary agent interface overnight
  • Forced rebuild using Claude Code + Anthropic API directly
  • Result: cleaner architecture, more control, less dependency risk
What it cost: A week of rebuild time. What it gave: a better system.
05
Apr 2026 Governed Intelligence
Current state.
  • Lean stack: Dee + Claude Code + on-demand agents
  • Constitutional governance operational
  • Custom skills codifying recurring workflows
  • 7 Cloudflare projects live — overnight autonomy is the next frontier
Where we are now.
02

Current Architecture

L1
Interfaces
How I interact with the system
Claude Code CLI Telegram DM WhatsApp groups Discord Proprietary dashboards
L2
Routing + Messaging
True Torque — Python webhook server (port 4343). Receives Telegram + WA, spawns Claude CLI sessions per chat with --continue for persistence. Cloudflare Tunnel for zero-trust ingress.
True Torque Cloudflare Tunnel WA bindings Telegram outbound
L3
Agent Stack
Deliberately lean. Dee (judgment) → Claude Code (primary intelligence) → on-demand agents spawned as needed. Built and dismantled: full orchestrator / secretary / overseer hierarchy. Went too big too soon.
Dee Claude Code On-demand agents Orchestrators Secretary Overseer
L4
Infrastructure
Cloudflare Workers / Pages / D1 across 7 active projects. Task system with proof discipline — every completion needs a file artifact. Session state, model routing, and custom skills as primary workflow tooling.
Cloudflare ×7 Task system Session state Model routing Custom skills
03

Key Decisions

01
Writing a Constitution
When agents operate autonomously across multiple surfaces, informal rules break down fast. A root governance document with hard constraints tied to real incidents keeps the system coherent as it scales.
02
Proof Discipline
Conversational acknowledgement isn't resolution. Every task completion requires a file artifact — if it's not written, it didn't happen.
03
Retiring GAS
Google Apps Script was the correct foundation 12 months ago, but AI capability moved faster than expected. All automation is migrating to Cloudflare Workers / Pages / D1 as the programmable, agent-ready foundation.
04
Building True Torque
Rather than depending on a third-party platform for WhatsApp and Telegram routing, a lightweight Python webhook server gives full control over message handling and agent dispatch.
05
Pulling Back on Complexity
Built a full orchestrator / secretary / overseer hierarchy too early. Pulled it back to: Dee + Claude Code + on-demand agents. Stability before complexity.
06
Custom Skills as Workflow
Recurring patterns — debugging, planning, code review, brainstorming — are codified as skills that enforce consistent process rather than relying on ad-hoc prompting each session.
07
Direct-to-Claude After OC
When Anthropic restricted OC from subscription plans, it forced a rebuild using Claude Code and the API directly. The constraint produced a cleaner, more controllable architecture.
08
Cloudflare Tunnel for Security
Webhooks need to be reachable from the internet without exposing a home IP or opening firewall ports. Cloudflare Tunnel provides secure, zero-trust ingress with no network configuration.
04

Lessons

01
Don't build agent hierarchy before you have stable workflows.
The orchestrator dream was premature — complexity introduced before the current layer was stable creates fragility, not capability.
02
Constraint is a catalyst.
The OC restriction forced a cleaner architecture than the one it replaced. Being pushed off a dependency turned out to be a forcing function.
03
Boring before next.
Each layer needs to prove it's stable before adding the next. This is now a core principle — but it took expensive lessons to get there.
04
Interfaces matter less than you think.
Spent time on Discord integrations, custom dashboards, proprietary surfaces. The answer was WA + Telegram + CLI. Simple surfaces, consistent channels.
05
Write things down structurally, not conversationally.
Discussed the same problems across multiple sessions without fixing them. The Constitution came directly from learning this the hard way.
06
GAS was right for its time.
The correct call 12 months ago — but AI capability moved faster than anyone expected. What was a reasonable automation foundation became a ceiling almost overnight.
05

Tools + Capabilities

AI Models What I can call
Claude (Anthropic) live Primary intelligence. Code, reasoning, long-form tasks. 11 per-project API keys.
Gemini (text) live Multimodal text + video understanding. Tier 1 Postpay billing.
Nano Banana Pro live Diagram + visual generation from text prompts. gemini-3-pro-image-preview.
Imagen 4 live Photo-realistic image generation. Same API key as Gemini.
GPT / Codex live OpenAI models accessed directly. Independent of other infrastructure.
Gemma (local) live Full local LLM pulled with Ollama. Runs on-device — private, zero API cost, no latency.
Veo 3 pending AI video generation. Vertex AI — needs aiplatform enabled on GCP project.
Gemini AI Studio manual Free multimodal playground in browser. No automation path — manual use only.
Dev Environment How I work
Claude Code CLI live Primary interactive intelligence layer. This.
Superpowers Skills live Codified workflows: brainstorming, executing-plans, frontend-design, code review.
GitHub live Private org (dee-CL-2026). Source control for all Cloudflare projects.
Wrangler live Cloudflare deployments via CLI. Token-based auth — no browser login required.
Integrations Connected services
Telegram live Inbound + outbound via True Torque. Per-chat Claude sessions, Dee's DM + groups.
WhatsApp live Group bindings, agent personas per group, outbound proactive messaging.
Zoom live Meeting end → auto-download transcript → AI summary → email to stakeholders.
Gmail / SMTP live Outbound email. Zoom summaries, Candid comms, agent-initiated sends (HG-gated).
Google Calendar live MCP tool — available inside Claude Code. Event read/write/find-free-time.
MCP Tools Available inside Claude Code sessions
Exa live Web search + crawl + code context retrieval.
Firecrawl live Web scraping, crawling, browser automation, structured extraction.
Tavily live Web search + deep research + multi-page crawl.
Playwright live Full browser automation — navigate, screenshot, click, fill forms.
Context7 live Library docs lookup by package and version. Used during code work.
Hardware What it all runs on
Mac Mini M4 handles everything 16GB RAM / 256GB. Runs True Torque, Ollama (local LLMs), Zoom webhook server, Cloudflare Tunnel, and all active dev work — concurrently, without breaking a sweat.