Blue: the ShieldAllowed actions go through.
Allowed actions go through.
Everything else is blocked.
The Shield sits between your agent and its tools. Every action is checked against rules you define: if it's allowed, it goes through; if not, it's blocked and logged. It runs on your side and its rules stay readable.
What's planned
Allowed-action list
Any action not in the rules is refused by default.
Per-action limits
Scope, amounts, paths, recipients: allowing an email doesn't mean allowing any email.
Flow rules
Sensitive data the agent has read can't leave through an outbound action.
Decision log
Every decision is timestamped, searchable and exportable to your monitoring tools. Rules are versioned in your repository.
Runs on your side
The software runs in your infrastructure: you keep the rules, the log and the code.
Ideal for
- Agents in production with access to sensitive data
- Teams moving from one-off testing to permanent control
- Answering customer questionnaires about AI over the long term
- Enforcing the rules written during a red team