A governance agent that profiles, cleans, and documents datasets — with a human in the loop. By Skylark 118 LLC · Version 1.0

The Data Steward Agent is a runnable data-governance agent for SaaS products. It acts as a blocking gate: no data feature ships until integrity, privacy, and change control are validated — and it never mutates data without explicit human approval.

This repo ships the agent in three forms:

Most data governance is either too enterprise (500-page manuals) or too vague (generic checklists). This is the essential 20% that delivers 80% of the value, packaged as an agent instead of a PDF.

The agent enforces three sequential validation gates. Each has a STOP condition; nothing advances until the current gate passes.

graph TB
    A[Data Steward Agent<br/>Human-in-the-loop]
    A --> B[Gate 1: Data Integrity]
    A --> C[Gate 2: Privacy & Security]
    A --> D[Gate 3: Change Management]
    B --> E[Production Ready]
    C --> E
    D --> E
    style A fill:#6B8FB3,stroke:#333,stroke-width:4px,color:#fff
    style B fill:#A23B72,stroke:#333,stroke-width:3px,color:#fff
    style C fill:#E4A04A,stroke:#333,stroke-width:3px,color:#fff
    style D fill:#6A994E,stroke:#333,stroke-width:3px,color:#fff
    style E fill:#3D4F5F,stroke:#333,stroke-width:4px,color:#fff

1. Data Integrity — required fields, referential integrity, formats, documentation
2. Privacy & Security — PII classification, legal basis, tested deletion (GDPR/CCPA)
3. Change Management — schema version control, paired forward/rollback migrations, impact assessment

Full procedure and rationale: docs/methodology.md.

Run the gates against a seeded dataset — pure Python standard library, no dependencies:

python3 demo/run_gates.py

You'll see the steward BLOCK a dataset with seeded violations (null required fields, orphaned records, undocumented PII, a migration with no rollback). Run python3 demo/run_gates.py --scenario clean to watch all three gates pass. Full walkthrough and captured output in demo/.

1. Copy the agent definition into your project:
   • AGENTS.md (or merge it into your existing one)
   • .claude/agents/data-steward.md
2. Ask Claude Code to run the data-steward subagent against your dataset.
3. The agent profiles read-only, then pauses for your approval before any deletion or migration.

Point your agent at AGENTS.md as the system prompt. Wire its write/DDL tools to a require-approval policy (see the SDK note below).

1. Read the methodology and follow the 4-week implementation guide
2. Copy and fill the templates:
   • PII inventory
   • Data dictionary
   • Schema change record
   • Deletion test checklist
3. Run the validation queries

The Data Steward is a governance agent: it investigates and proposes, and a human approves before anything consequential (deletion, migration, any write). The pattern is the same in every harness — gate the write tools behind an approval step.

Install the official SDK — npm install @anthropic-ai/sdk (Python: pip install anthropic). The snippet below shows the core idea: a tool-use loop that requires human approval before executing any write tool. The system prompt is the agent definition from this repo.

import Anthropic from "@anthropic-ai/sdk";
import { readFileSync } from "node:fs";

const client = new Anthropic();
const systemPrompt = readFileSync("AGENTS.md", "utf8"); // the agent definition

// Tools the Data Steward can call. Reads run freely; writes need a human.
const WRITE_TOOLS = new Set(["run_write_sql"]); // delete / migrate / anonymize

async function approve(toolName: string, input: unknown): Promise<boolean> {
  // Replace with your real approval UI / Slack prompt / CLI confirm.
  console.log(`Approval needed for ${toolName}:`, input);
  return /* await human decision */ false;
}

// In your tool-use loop, before executing a tool the model requested:
async function executeTool(toolName: string, input: unknown) {
  if (WRITE_TOOLS.has(toolName) && !(await approve(toolName, input))) {
    return { error: "Denied by human reviewer — action not taken." };
  }
  // ...run the read-only or approved tool...
}

This approval gate is harness-agnostic and works with any model loop. For Anthropic's hosted Agents (managed permission policies, where always_ask pauses a session for confirmation server-side), see the official docs: permission policies · agent setup.

data-steward-agent/
├── AGENTS.md                          # Canonical agent definition (portable)
├── .claude/agents/data-steward.md     # Runnable Claude Code subagent
├── docs/
│   ├── methodology.md                 # The 3-gate framework, in full
│   └── implementation-guide.md        # 4-week human rollout plan
├── templates/                         # The agent's output artifacts
│   ├── pii-inventory.yml
│   ├── data-dictionary-template.md
│   ├── schema-change-template.md
│   └── deletion-test-checklist.md
├── scripts/validation-queries.sql     # Gate 1–3 check suite (a tool the agent runs)
├── demo/                              # Runnable proof — seeds bad data, runs the gates
│   ├── run_gates.py                   #   zero-dependency demo (Python stdlib)
│   └── sample-report.md               #   captured PASS/BLOCK output
├── LICENSE                            # MIT
└── DISCLAIMER.md                      # Not legal advice

Every SaaS product eventually hits data-governance failures: bad data quality breaks analytics; missing privacy controls invite fines (GDPR: up to 4% of revenue); no deletion capability is legal liability; uncontrolled schema changes cause outages. This agent is built to catch those before they ship.

Distilled from real production work in a multi-tenant SaaS product handling sensitive health data — practice, not theory.

Start with the three gates. Add enterprise features only when you need them:

• Performance monitoring → when queries regularly exceed 1 second
• Advanced backup/DR → when preparing for compliance audits
• Detailed audit logging → when sensitive-data access must be tracked
• Quarterly compliance reviews → when regulations require formal audits

Improvements welcome. Please:

1. Keep the framework focused — resist feature bloat
2. Prioritize practicality over comprehensiveness
3. Include a working example with your PR

MIT — use freely. See LICENSE. This repository is not legal, privacy, or compliance advice; see DISCLAIMER.md.

Built by Skylark 118 LLC. Need help putting governance guardrails on your product or your agents? Work with Skylark 118 →