The confirm gate: letting an AI act without letting it loose

Proposals, not actions

When Macy spots a trigger you set, she does not act. She proposes. The proposal arrives as a plain text with everything you need to decide in one glance: what happened, what she suggests, what it costs. The action fires only when you reply yes.

This sounds like a limitation. It is actually the unlock. Because every risky action waits for a human word, we can connect Macy to things that would otherwise be unthinkable for an AI, like a brokerage account. Robinhood built their agentic API around exactly this pattern, and it is why we integrated it first.

Logs are a feature

Every proposal, every yes, and every execution is logged with a timestamp you can read. Not for compliance theater. Because trust in an agent is built the same way trust in a person is built, by being able to check what they did and finding that it matches what they said.

What we will not build

A mode where Macy trades without asking. Some users will want it. The day an agent moves money on a hallucinated signal is the day this category dies. The yes stays.