Live, on a real account
Everything here runs live. Pressing the button creates a Sheaf account for you, writes an AI agent from one sentence, gives it tools, and lets it run. An agent is an AI that can take actions on its own. It takes about two minutes. We do not know in advance exactly what the agent will do.
Nobody writes code. You describe the job in English and Sheaf produces the agent: which AI model it uses, what it is told, which tools it may use, how many steps it gets. You can read every line before it runs.
The code is produced directly from that description, so what you read is what runs. You can take it and run it on your own machine. It still asks Sheaf before every tool it uses, so the safety check goes with it.
A set of finance tools is attached. Sheaf asks what the tools can do and marks every one as needing a person, because the tools themselves cannot say which actions matter. We allow the harmless lookup to run on its own and let several AI models judge the payment.
Every step is written down before the next one starts. If the machine running it died right now, another would pick it up from here rather than start again.
The agent wants to move money. Several other AI models decide that call again, separately from the agent that wants it. Sheaf reports how much they agreed.
Every decision goes into a tamper-evident record. Change one entry and every entry after it shows the break. A person releases or refuses what was parked, and that goes into the record too.