Guides
Human approval set up for one kind of risky agent action at a time: the risk, the rules to write, the agent's code, what the approver sees, and how to roll it out.
Customer-support refunds
Let a support agent refund small orders on its own, send larger ones to the people who handle refunds, and never refund the same order twice.
Read the guideAlways ask
refund.*above €50.00Production deploys and infrastructure
Hold terraform apply, kubectl and production deploys until a person says yes, whether they run from CI, a script or a coding agent's hook.
Read the guideAlways ask2 people approve
infra.destroy.*Payments and finance
A cap per payment, a budget per day and two people for the large ones: a rule set for agents that pay suppliers or move money.
Read the guideAlways ask2 people approve
payment.*above €1,000.00Outbound email and CRM changes
A reply to one customer can run on its own. A send to 4,000 contacts or a bulk CRM update waits for a person who can see the list first.
Read the guideAlways ask
email.send_bulkData deletion and GDPR erasure
Erasure cannot be undone and has a deadline. Let the agent prepare each deletion, a person approve it, and the record show who did.
Read the guideAlways ask
data.erase.customerDatabase migrations and destructive SQL
Migrations, DROP, TRUNCATE and a DELETE without a WHERE wait for a person who can read the statement, from CI, a script or a coding agent.
Read the guideAlways ask2 people approve
db.migrate.production
Every guide has the same five parts
- The risk
- What goes wrong when the agent acts alone.
- The rules
- A rule set to start from, as you would write it.
- The code
- The agent's side in TypeScript, Python, curl and MCP.
- The approver
- What the person deciding sees, and can change.
- The rollout
- From asking about everything to rules for the routine.