Building Autonomous Agent Marketplaces at CVS Health

When I joined CVS Health as a consultant in August 2025, the brief was ambitious: build the infrastructure that would let a Fortune 5 healthcare company deploy AI agents at enterprise scale.

Six months later, we have:

• Autonomous agent marketplace — Business units can deploy AI agents without engineering bottlenecks
• Internal "vibe coding" sandbox — 300K+ employees have Cursor-like development tools inside the security perimeter
• Multi-modal intake systems — AI agents handle ordering, documentation, and training
• Digital twin agents — Simulated environments for testing before production deployment

Here's how we did it.

The Problem: AI Adoption Bottlenecks

CVS has 300,000+ employees across retail pharmacies, healthcare clinics (MinuteClinic), insurance (Aetna), and corporate operations. Every business unit wants AI. But:

• Security review takes weeks per agent
• Engineering resources are scarce
• Compliance requirements (HIPAA) are non-negotiable
• Shadow AI is already happening (people using ChatGPT on personal accounts)

The solution wasn't to block AI adoption — it was to enable it safely.

The Architecture: Agent Marketplace with RBAC

We built a self-service platform where approved users can deploy AI agents. The key components:

1. Role-Based Access Control Layer

Not every employee gets the same capabilities:

• Pharmacy operations → Agents that can access prescription data (heavily restricted)
• MinuteClinic → Agents with patient interaction capabilities (HIPAA compliant)
• Aetna claims → Agents for claims processing (PHI access, full audit logging)
• Corporate → General-purpose agents with no healthcare data access

The RBAC layer determines what data an agent can access, what actions it can take, and what gets logged.

2. Sandbox Environments

Every agent deploys to a sandbox first. The sandbox:

• Uses synthetic data (no real PHI)
• Simulates production APIs
• Logs all interactions for review
• Requires approval before production promotion

This means business users can experiment with AI agents without security bottlenecks — but nothing touches real data until it's approved.

3. Multi-Modal Intake

We built AI agents that handle intake across channels:

• Voice — Phone-based intake using Twilio + ElevenLabs
• Text — SMS and chat interfaces
• Document — PDF/image parsing for insurance claims
• API — Programmatic intake for enterprise integrations

The intake agents normalize requests, extract structured data, and route to appropriate systems.

4. Digital Twin Testing

Before an agent goes live, it runs against a "digital twin" of the target system. The twin:

• Mimics production behavior
• Uses anonymized historical data
• Catches edge cases before deployment
• Generates compliance documentation automatically

The "Vibe Coding" Sandbox

One of the most popular features: an internal development environment that feels like Cursor or Replit, but runs entirely inside CVS's security perimeter.

Employees can:

• Prompt AI to generate code
• Test in sandboxed environments
• Deploy to approved infrastructure
• All without leaving the corporate network

This replaced shadow AI usage with a sanctioned, logged, compliant alternative.

Results After 6 Months

• 40% reduction in deployment time for AI-powered features
• 12 business units actively using the agent marketplace
• 50+ agents deployed in production
• Zero HIPAA violations from AI-related activity

What's Next

We're now building:

• Agent-to-agent communication — Agents that can delegate tasks to other agents
• MCP integration — Model Context Protocol for standardized tool access
• Automated compliance review — AI-powered pre-approval for low-risk agents

Lessons Learned

1. Enable, don't block — Shadow AI happens when legitimate needs aren't met
2. RBAC is everything — Granular permissions unlock adoption without violating security
3. Sandbox first — Let people experiment safely, then promote to production
4. Audit everything — Regulators will ask. Be ready.

The future of enterprise AI isn't a single model — it's a marketplace of specialized agents, each with carefully scoped permissions, running in isolated environments, with comprehensive audit trails.

That's what we're building at CVS Health.

---

Ian Noble is a consultant leading AI automation at CVS Health. He has 30+ years of enterprise platform experience and has deployed 25+ AI applications.