Nothing asks you anymore: System One decides, and every agent starts fresh
Tool and permission switches are gone. System One judges risky steps in the background, helps the agent as it works, and every agent starts empty.
Allowed, with a judge behind it
Every tool is allowed and Bandicot never stops to ask you. When a step is risky, such as a dangerous command, a write outside the project, acting after reading a web page or a team message, or changing the agent's own control files, the System One model you picked decides. It lets the step through only when it is sure enough, otherwise it blocks it and tells the agent why so it tries another way.
If System One does not answer in time, fixed rules block the dangerous commands, the control files and messages to other people. A daily limit of 5 dollars guards paid model calls. Every decision shows in the System One activity list.
Working beside the agent
System One now reminds the agent of what was already said when a step fails, keeps it going when the work is not finished, catches steps that failed quietly, and suggests the right tool for the task. The agent can also ask it to rank, choose, check or recall.
Fresh agents, one system prompt
Every new agent starts with nothing loaded, only Bandicot's prompt, your system prompt from Settings, Agent, and the project's AGENTS.md. It loads skills, memory, history and connections itself when it needs them. The inventory is put away for now.