Wherever an agent can act, put a gate
Each of these maps to a worked example in the docs — with the exact policies to install.
Customer-facing email agents
An agent drafting and sending customer email should not have unilateral send rights. Gate send_email behind human approval; reviewers see the payload summary and approve or reject from the dashboard. Every send — attempted or completed — is in the audit log.
Purchasing & payments agents
Cap what an agent can spend per day and per month, require approval above a dollar threshold (for example $100), and keep a cost-attributed trail of every transaction it attempted. Money never moves without a human when you say so.
Data & infrastructure agents
Block destructive operations by pattern — *delete*, *drop*, *destroy*, *wipe* — org-wide, with explicit allow rules for the narrow cases you've reviewed. Flip the org default to block and unknown tools can't run at all.
Third-party MCP tools
Adopting an MCP server you didn't write? Wrap it with the Ruligent MCP gateway — one command, no code changes. Every tools/call is guarded; blocked calls return a tool error to the client and never reach the server.
Customer support agents
Let a support agent read anything but gate refunds, account changes, and escalations. Rate-limit high-frequency actions per rule so a confused agent can't spam customers.
Internal ops & compliance
When leadership asks "what are our agents actually doing?", answer with data: per-agent activity, decision breakdowns, spend attribution, and exportable audit history for reviews and vendor assessments.
Validate an idea before building
The Validation Agent produces an opportunity report with a BUILD / PIVOT / RESEARCH_MORE / STOP recommendation. Every claim is labeled known fact, assumption, or needs verification — with concrete ways to check. STOP is a first-class answer, not a failure.
Prepare a launch with governance
The Launch Agent scores readiness deterministically across 10 categories — computed from real project and org state, no self-reported numbers — and drafts launch copy for your review. Anything risky (external communication, spend, deletes, deploys) always waits for a human approval.
Your agents, your rules.
Install the recommended safety pack in one click, then tune it to your risk tolerance.