MCP Tools in Production: Interface, Runtime, and Policy Boundary

Where MCP fits in a production agent system, and where you still need sandboxing, policy, logs, and result contracts.

Audience: Teams adding MCP tools to customer-facing agent workflows.

The problem

MCP gives agents a way to call tools, but it does not decide where code runs, how secrets are scoped, or how outputs become product data.

Implementation path

Use MCP for controlled capabilities and Argo for the run boundary: files, sandbox execution, logs, artifacts, and final result validation.

Tradeoffs and failure modes

MCP standardizes capability access; the runtime still needs to enforce execution and data boundaries.

Boundary diagram

Model <-> Broker policy <-> MCP tools
                 |
                 v
          Argo sandbox runtime
          inputs / bash / files / artifacts

Run this on Argo