Playbook

1. Write the Operating Policy

Define allowed reads, writes, commands, network targets, external services, models, and approval boundaries independently of vendor syntax.

2. Map Every Execution Path

Include shell, file tools, MCPMCPUSB-C for AI — one standard plug so any agent can connect to any data source or tool.The Model Context Protocol: an open standard (Anthropic, 2024) for connecting AI systems to external tools and data without building a custom integration for each one.Full definitionClose, subagents, IDE actions, browser tools, and cloud workers. A control that covers one path is not global policy.

3. Build Product Adapters

Translate the policy into Claude Code settings, Codex configuration, and Copilot settings for each supported surface. Keep generated or reviewed adapters traceable to the same policy version.

4. Lock Higher-Trust Controls

Use managed settings, OS or container isolation, service-side scopes, and protected CI where a repository or local user must not weaken the rule.

5. Inspect Effective State

Capture active sources, resolved permissions, connected extensions, and model selection at run start.

6. Run Negative Tests

Attempt a denied file read, out-of-scope write, network request, external tool action, nested-agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose bypass, and unapproved publication.

7. Requalify Cloud Surfaces

Do not infer hosted behavior from local configuration. Verify identity, filesystem, network, secrets, and approval behavior in the deployed environment.