본문 바로가기
Case · 2026.10.14

Before you build agents, lay down registration, a gateway, memory, tracking and PII blocking as one set — so every agent passes through the same control path

Paving the Road to Production: Governed Enterprise AI Agents — Harish Gaggar, Credit Karma — AI Engineer · Read the original ↗

This page is a translation of the Korean original. Read the Korean original →

What you get

  • A single portal registration creates the repository plus identity, memory, tracking and secrets resources to the same standard
  • Every tool call is authenticated at a central gateway, so direct access to endpoints disappears
  • Each query leaves a record of model, tokens, cost and quality that you can aggregate and answer from later

When to use it

When you are in an air-gapped environment and cannot use external MCP servers, so you have to build your own MCP servers and gateway; and when you run hundreds to thousands of agents and, six months on, must aggregate 'which agent did what and how many tokens it used' into an executive dashboard.

How to do it

  1. Register the use case, agent ID, framework (LangGraph, etc.), language and skills/prompts in the internal developer portal, and get the repository and supporting frameworks (memory, tracking/evaluation, context optimization, PII blocking) generated in one go.
  2. Clone the repository locally, build and test it as a Docker image, store traces and tracking/evaluation results in SQLite or BigQuery to confirm behavior, then open a PR that needs two approvals before it goes to CI/CD.
  3. In CI/CD, register the agent in the central registry, build the container image and scan it for vulnerabilities; once security approves, push it to the registry and deploy it to a Kubernetes namespace.
  4. Build your own MCP servers with a FastMCP wrapper; every tool goes through internal security review and only approved ones reach production, and every call must pass JWT authentication at the MCP gateway before it reaches an MCP server or endpoint.
  5. In the LangGraph workflow, interpret and classify the query, then gather evidence and run policy checks; execute read-only actions directly, send write or sensitive actions through HITL approval, and record both approvals and rejections in the audit log.

What to borrow

The approach of defining 'what gets created automatically when you register' before writing any agent code is worth adopting directly. Two points stand out: the branch that lets read-only actions through and routes only write or sensitive actions to human approval while logging rejections too, and the habit of writing model, tokens, cost and quality to a table on every run. The 90% token saving, however, is presented without a comparison of answer quality, so check it against quality in your own environment.

Background — what the original covers

Harish Gaggar of Intuit Credit Karma describes the common path he built to move experimental agents into in-house production. The flow runs from developer-portal registration through container build, security scanning and Kubernetes deployment, with custom MCP servers and a central gateway, LangGraph workflows, memory checkpoints, tracking and evaluation tables, PII blocking and context optimization attached as one bundle. The demo inspects the record tables and does not validate overall quality scores.

#Agents#Safety

← All practitioner tips