Jul 2026
An agent that fails closed
Rivesa Team Engineering
The interesting question about an autonomous agent is not what it does when everything works. It is what it does when something breaks.
Every agent demo shows the happy path. The parts that decide whether you can actually give one your calendar are the unhappy ones: an unclear instruction, a half-configured connector, a run cancelled mid-flight, a model that asks for something it was never granted.
Rivesa is built so that all of those end in nothing happening. Deny is the default, and every gate has to be passed on purpose.
Confirmation defaults to no
Anything with a real-world side effect — sending, booking, changing — is gated on an explicit approval tied to an actual person's decision. Absent that, the answer is no. Not 'ask again', not 'assume yes because the user seemed keen'.
A cancelled run cannot still fire an effect. Cancellation is threaded all the way to the outbound call, so stopping an agent actually stops it rather than stopping your view of it.
Permission is not a string the model can guess
Agents are granted capabilities, not raw tool names, and the server re-checks each one independently rather than trusting what the caller claims. Identity comes only from a verified token — never from a field in the request saying who someone is.
The practical effect: text the model can invent is not a key that opens anything. There is no phrasing that talks its way past a gate, because the gate is not reading the phrasing.
"Deny by default is not caution. It is the only version of autonomy anyone can afford to switch on."
Boring on purpose
Quotas, run timeouts, size limits, truncated tool output: none of it is exciting, and all of it is the difference between an agent you demo and an agent you leave running.
A thousand-odd automated tests hold this shape in place. Not because tests prove an agent is wise, but because they prove the boundaries are still where we said they were.
Rivesa Team writes product notes for Rivesa, focusing on agent architecture, permissions, and the craft of keeping delegated work accountable.