I’m working on a software-agent system and keep coming back to one design question: **Should the model/provider that performs an action be allowed to be the final authority on whether the action succeeded?** My current answer is “no,” at least for meaningful software work. I’m building Flows around a chain where execution, checks, repair, and evidence are separate concepts. Oort is the canonical library/provider layer underneath it. https://flows.oortstack.com https://oortstack.com ...