Kevin Z.

Check out my presentation at Collision Conf for MyWiFi #socialwifi #wifimarketing #saas facebook.com/kzicherman/pos…

@bcherny Check out commanderplugin.com

Big fan of teaching more people the basics of using Claude Code in an accessible way. So much of the world has not yet used agents. There's a lot of opportunity to level the playing field and expand access.

@Suryanshti777 Opus AI PM for Claude Code tha that picks which plugins you should use - commanderplugin.com

Claude Code feels completely different once you install this. Anthropic quietly released an official plugin called claude-code-setup and it basically turns Claude Code from “pretty good” into an actual AI dev environment. It scans your project and recommends: → hooks → skills → MCP servers → subagents → automations Then sets everything up step-by-step for you. Most people are using Claude Code completely vanilla… which is why their experience feels messy. The real power comes from the ecosystem around it. Install: /plugin install claude-code-setup@claude-plugins-official Bookmark this before you forget it.

@DivyanshT91162 GitHub.com/kevinzai/comma…

@DivyanshT91162 Check out commanderplugin.com

Claude Code suddenly makes WAY more sense after installing this official Anthropic plugin 🤯 It reads your entire project and instantly tells you what you should actually enable, use, and configure. The crazy part? Most people using Claude Code don’t even know this exists. → Detects your stack automatically → Understands your codebase structure → Suggests the right workflows → Removes the confusing setup phase → Makes Claude Code feel 10x smarter This is probably the real onboarding Claude Code needed 🔥

# Why 36 Agents Beat 1: The Architecture Nobody Shows You Running a single AI agent is easy. Running 36 in production — across self-hosted infrastructure and 5 cloud platforms — is a systems problem. Most people never get there because they hit context collapse first. The ones who do usually stop because the coordination overhead feels impossible. I didn't stop. This is what actually works at scale: the four-layer architecture I've been running for 18 months. 36 agents. 5 product companies. 24/7 autonomous operation. $2.48/agent/day. 99.97% uptime. Real numbers. ## The Single-Agent Ceiling One agent can't hold the full context of five companies simultaneously. One agent doing sales, engineering, marketing, and support produces mediocre results across all of them. You hit the ceiling fast. The fleet model trades coordination overhead for specialization. Each agent gets a defined role, bounded context, and a communication protocol for crossing boundaries. Think management hierarchy: I direct senior agents. Senior agents route to specialists. Specialists execute and report back. I only see escalations and summaries. One agent does 70% of things fine. Five specialized agents do those same things at 95% quality. The math compounds. ## Layer 1: The Gateway (The Traffic Controller) Every LLM request — whether from my terminal, a scheduled cron, or an agent spawning a sub-agent — routes through the gateway before hitting any model provider. What it does: - **Model routing**: selects the right provider based on task type, cost, and current rate limits - **Load balancing**: distributes across multiple accounts to avoid per-account quotas - **Budget enforcement**: hard kills at spending thresholds - **Logging**: every request logged with token counts, cost, latency, model used - **Health monitoring**: probed every 5 minutes; auto-alert after 3 failures The gateway runs as a persistent system service — always on, never restarted without explicit approval. If it goes down, all 36 agents stop. Six providers route through it: Anthropic (Claude via ClaudeSwap), OpenAI, Groq, Google Vertex, Cloudflare Workers AI, and Ollama local. Provider selection is automatic. ## Layer 2: The Orchestrator (Neo) Neo owns task dispatch. The rule is simple: any task needing 2+ workers or multi-day execution routes through Neo. He dispatches through Paperclip — the task management system running at a local endpoint. Every task gets tracked end-to-end. Nothing falls through. Every task has an owner, status, and log. Single-worker tasks that Alfred (the personal assistant) can handle inline → Alfred handles directly. This boundary is enforced on both sides. Result: multi-step work is auditable. You know exactly what's happening and when. ## Layer 3: The Specialist Roster 15 specialist agents own defined domains: - **Alfred** (Claude Sonnet): Personal assistant, coordinator - **Morpheus** (Claude Opus): System architect, compliance - **Neo** (GPT-5.5): Task dispatch - **Sentinel** (GPT-5.5): Research, architecture review - **Viper** (Grok): Trading and markets - **5 GuestNetworks domain leads** (Claude Opus each): CEO advisor, sales, support, marketing, dev, ops Three coding agents (Codex, Gemini, Claude Code) self-thread for complex tasks in dedicated channels. They cannot delegate to other LLMs. ## Layer 4: The Free Fleet (Zero-Cost Specialists) Four agents run on zero-cost infrastructure: - **Forge** (Groq Llama 4 Scout): Reasoning, complex logic - **Oracle** (Groq): Fast research, deep analysis - **Flare** (Cloudflare Workers AI): General tasks - **Lama** (Ollama local): Batch processing, offline compute The auto-dispatcher runs every 30 minutes, routing unassigned tasks to the free fleet when appropriate. Roughly 40% of tasks complete without touching paid models. At current volume, the free fleet handles hundreds of tasks per week at $0. That's load that would otherwise hit the paid tier. ## The Communication Protocol (Auditable Routing) Agents communicate via `sessions_send` — direct messages between sessions. No Discord for inter-agent routing. Every send follows an atomic pattern:

the full operator's manual: → kevinz.ai 38 agents · workflows · real costs · what broke no fluff. real numbers only. $ explore →

I came back to X because most people don't know this is possible. So I'm documenting it all — the agents, the workflows, the failures, the costs. Follow if you want to see how one operator runs an entire stack with AI.

In 2019 I was running a WiFi marketing SaaS. 3k customers. Team of 15. I posted about Zapier integrations and conference talks. Normal CEO content.

2,468 days since my last Twitter post. An army of AI agents now run my business and manage my life. Back to tell you everything

@elonmusk All your shares are belong to @elonmusk youtube.com/watch?v=qItugh…

Taking Twitter private at $54.20 should be up to shareholders, not the board

@elonmusk KYC accounts to enable verified DOGE wallets… possibilities are endless 🐕

Wow 13 years! Do you remember when you joined Twitter? #MyTwitterAnniversary

MyWiFi Networks is hiring! Join our dynamic team in a fast-paced entrepreneurial startup environment. APPLY NOW ➔ buff.ly/2vGODnF

Thanks Zapier for the awesome blog post. Our Social Wi-Fi Marketing Automation Integration is now LIVE! snip.ly/5wzg3

I just published “Social Wi-Fi Marketing Automation Just Got An Upgrade With MyWiFi Networks and Zapier” medium.com/p/social-wi-fi…