OgreScat/AgentGRIT ? reverse-engineered prompt

Build me a Python project that acts like a governance first AI agent runtime, not a single ready made bot.

I want it to route each task to the cheapest capable model first, use a local model when possible, and only go to cloud models when the task really needs it. The agents should follow their own rules, block obviously dangerous actions, verify work before saying it is done, and escalate risky stuff to a human instead of quietly guessing. If something is refused or uncertain, it should explain why and leave a clear trail of evidence.

Make it so I can plug in my own agents on top of the framework, run a main orchestrator, and also have a simple command line way to govern a one off task. Include a basic example agent, configuration through an env file, and a clean way to run tests and a quick smoke check. If you need to, look up current docs for any model or local runtime details.