I spent the last year building the part of AI systems nobody puts in the demo: what happens when things go wrong. Most agent frameworks are built to answer one question, what can this model do? Nexus is built to answer a different one, what happens when it's wrong, and how do we know? Here's what that actually means, in plain terms: Every decision has one owner. No step in the system can quietly act without something else checking it. Writing, deleting, or changing anything real requires a human to say yes or not as a courtesy, but as a hard rule the system can't route around. Nothing is remembered by guesswork. Every action is logged as it happens, in order, permanently. If something breaks, I don't have to reconstruct what happened from memory or hope the logs are enough. I can replay the exact sequence of events and see precisely where it went wrong. Recovery isn't a footnote. A system that fails silently and reports success anyway is worse than one that crashes loudly. Nexus is built so failure is visible, traceable, and recoverable not swept under a "task complete." The result feels less like an agent framework and more like an operating system, one designed around a simple belief: the hard problem in AI isn't making a good decision. It's knowing, with certainty, what actually happened after you made it. Nexus v2.0 just shipped. It's open source, and it's the clearest version yet of everything above. GitHub: https://github.com/STiFLeR7/nexus Would genuinely love feedback from anyone building production AI systems.