Illustrative case study: coding-agent production access

A hypothetical five-developer SaaS team evaluates one production tool boundary.

Illustrative example · not an actual customer

Initial risk

In this scenario, coding agents share a credential for repository, database and internal review tools. Per-call context and review ownership are missing from that baseline.

Agent and MCP workflow

Five sample identities use an HTTP gateway and a local command-mode fixture for 14 days. Calls use synthetic data and do not touch a customer's infrastructure.

Control model

The teaching model denies destructive shell commands, reviews sensitive actions and allows routine reads. The production pack continues to deny database writes; the teaching model is not a claim of a configurable production DSL.

Evaluation outcome

Operational result

The scenario recommends a bounded read-only rollout, a named policy owner and a separate approval reviewer. These numbers do not establish financial savings or prevented incidents.