How do I let an agent act without letting it act wrongly?
Safety
Confirm-first actions, biometric tiering, scoped tokens and the audit trail you need before an agent touches money.
7 posts
Start here
Confirm cards: letting an agent act without losing control
If the confirmation lives in the prompt it is a suggestion. Make it a branch in your server code, and decide the tier per tool rather than per sentence.
- Designing a confirm-and-pay flow an agent can runThe payment path end to end: where the write is intercepted, what the card must show, when a biometric is worth asking for, and why speech is never the credential.Safety10 min
- An agent in a banking app, and the gate before money movesHow to let an agent act inside a fintech app: which actions get a confirm card, which get a biometric, and why the gate lives in code.Safety8 min
- Security and privacy for an agent that can act on your dataThe threat model for an in-app agent is its tool set, not its microphone. The OWASP entries that apply, the injection path, and the controls worth building first.Safety9 min
- Does your agent work eight times out of eight?Average accuracy is the wrong number for a product. Run the same task eight times and count how often it worked every single time.Safety11 min
- Prompt injection when the agent can spend moneyIn a consumer app the untrusted text is your own user's data and the tools move their money, which makes filtering useless and structure the only real defence.Safety10 min
- Which agent actions deserve Face ID, and which do notBiometric-gating everything trains users to approve without reading. Tier agent actions by what they can destroy, and let the platform decide how.Safety10 min