Four things it does
Every Sheaf product is built on these four. Sheaf One offers them directly, for teams that run their own AI agents and want Sheaf underneath all of them.
Ask
One question goes to several AI models that you choose. Each answers on its own and cannot see the others. Sheaf Pod returns one answer and the check behind it.
Check
The agents' claims about the same facts are checked for agreement, and for the contradictions that only show up when you look at all the claims together. The result is a signed record of whether the reasoning held together, with any claims in conflict named.
Stop
Tell Sheaf which action a decision leads to. If the reasoning behind it contradicts itself, Sheaf stops the action and parks it until a person releases it. The agent cannot go around this.
Record
Every check and every stopped action goes into a tamper-evident record, signed and timestamped by an independent timestamp authority. The record is ready for an audit.
Three ways to use it with your agents
Keep your own agents and let Sheaf watch from outside, either warning you or stopping the action. Or let Sheaf run the AI models for you, with the stop built in. All three give you the same signed record.
Check only
Send your agents' answers to Sheaf. Sheaf returns the signed record and points out any contradictions. Your system decides what to do.
Stop the action
Sheaf sits between your agent and the tools it uses, through the standard way tools are connected to an AI agent. An action that rests on a contradiction is parked for a person to release before it happens.
Let Sheaf run it
Sheaf asks the AI models and returns one answer. Tell it which action follows, and a contradiction parks that action on its own.
Start with the AI Agent Safety Check
Before you connect anything, Sheaf scores your AI agents on seven measures, such as whether important actions are checked before they happen. The report shows each weakness and what Sheaf would do about it. It takes a few minutes and needs no code.
Two products are built on it
Sheaf One is the general version. For regulated decisions we also supply it set up for one industry, with that industry's terms, rules, and types of case already in place.
Pulls any number of data sources into one picture and checks that they agree before a decision rests on it. First use: mortgage and credit broking.
AI agents that assess credit risk across a whole loan book and each customer in it. Decisions that rest on a contradiction are stopped for a person.