Our Approach

4 layers to scale AI safely across 50+ engineers

About Bowtie www.bowtie.com.hk

• HK's first virtual insurer (licensed Dec 2018)
• Direct-to-consumer, zero agents, zero commissions
• Series C: US$70M (2025) · ARR > US$80M
• In-house core system on AWS Cloud
• Using Copilot since 2022, GenAI since early 2023
• 200+ employees, ~50 engineers

🔑 Why Developers Get to Vibe Code

Our LLM Gateway

Single LiteLLM instance on AWS — serves engineers and internal services

Every Request Goes Through the Gateway

🛡️ Why This Didn't Hurt Us

Four practices that existed before this attack. None of them exotic.

Centralized Guardrails

📊 Cost & Adoption Visibility

Everything flows through one gateway — so we see everything

Why Not Just Pick One Provider?

Source: Artificial Analysis, Apr 2026

GPT → Claude → Gemini → DeepSeek → GPT → …
🔄 Engineers keep switching to the "best" model

No single provider wins on both intelligence and speed.

🧑‍💻 The editor problem

You can't force everyone onto one editor — nano vs vim vs VS Code vs Notepad++ vs JetBrains.

Same with LLMs — engineers have strong preferences on their "best coding model."

🔀 Gateway = freedom + control

Let engineers pick their model. The gateway enforces DLP, cost, and auth regardless.

我全部都要 — "We want everything!"

Case Study: Notion MCP

What the shared-secret era actually looked like

{
  "mcpServers": {
    "notionApi": {
      "command": "npx",
      "args": ["-y", "@notionhq/notion-mcp-server"],
      "env": {
        "NOTION_TOKEN": "ntn_****"
      }
    }
  }
}

{
  "mcpServers": {
    "notion": {
      "url": "https://mcp.notion.com/mcp"
    }
  }
}

This is the pattern we follow for all Bowtie MCP servers.

Three Types of MCP

{
  "mcpServers": {
    "notion": { "url": "https://mcp.notion.com/mcp" }
  }
}

MCP + OAuth Architecture

flowchart LR
    subgraph Internet["🌐"]
      Dev["💻 Developer"]
      CLI["🤖 AI Agent\n(Claude Code)"]
    end
    
    subgraph VPC["🔒 VPC Boundary"]
      MCP["📡 MCP Server\n(HTTP + OAuth)"]
      Tools["🛠️ Tools\n(list_tickets, get_user_info, ...)"]
    end
    
    subgraph Auth["🔐 OAuth 2.0"]
      Cognito["Amazon Cognito\n(User Pool)"]
    end
    
    CLI -->|MCP + Bearer JWT| MCP
    MCP --> Tools
    
    Dev -->|1. Login| Cognito
    Cognito -->|2. Access Token| Dev
    Dev -->|3. MCP Call + Token| MCP
    MCP -->|4. Validate| Cognito
  

Each user has their own token. AI acts as the user's delegate. Token revocable anytime.

The Golden Rule

Our MCP Toolkit

9 servers — all Streamable HTTP, all behind the gateway

Same endpoint pattern: https://mcp.bowtie.internal/{server}/mcp (example)

Who Uses What

Same tools, different workflows — the surprise was who adopted fastest

3 engineers → 50+ people in 6 months — because the safe path was the easy path

Wrapping Bowtie API with OAuth MCP

Existing Cognito-backed API → MCP HTTP server with full OAuth 2.0

What Is a Skill?

A portable package of knowledge, procedures, and code your agent loads on demand

The Computing Analogy

A few companies build processors. Millions build applications.

You don't need to understand CPUs to build an app.
You don't need to train models to build skills.

Why Skills Beat Raw Intelligence

A Skill in Practice

API investigation — our most-used skill

Same pattern for analytics (PII auto-masked), code review (per-repo CLAUDE.md), ticket triage (Linear workflows)

Give Your AI a Spine: CLAUDE.md

Your AI should push back like a senior engineer, not agree like an intern on day one.

SOPs that previously lived in someone's head — now version-controlled and enforced by AI.

Vibe Coding Went Enterprise

This isn't early adopters anymore

The Vibe Coding Tension

So… which one is right?

Vibe Coding ≠ Just Writing Code

What "coding" actually looks like in practice

The biggest productivity gain isn't in writing code faster — it's in everything before the code.

From 3 Pioneers to 50+ People

And we're not alone — Stripe, Sentry, Linear all ship official MCP servers as building blocks

Agents Are Already Shipping Code

This isn't hypothetical — it's happening at the biggest companies in tech

The pattern: agents write, humans review. Sound familiar?

The 24/7 Agent

Same 3 layers — just running without a human at the keyboard

If you built layers 1–3 well, the autonomous agent is just a different caller — not a new problem to solve.

Key Takeaways

Many "AI issues" were not caused by AI — only amplified. The fix was team-wide governance, not restriction.

Build these 3 layers well, and autonomous agents become a natural extension — not a new problem to solve.