26 Hours, Zero Prompts: 3 AI Agents Running Autonomously

I’m building an experimental control, memory, execution, and verification architecture for autonomous AI agents.
The overarching goal is to make it possible for an AI system to take a high-level objective and work on it continuously for days or eventually much longer, using a real computer and tools, without requiring a human to supervise every individual action. It should be able to inspect its environment, break goals into work, execute that work, test the results, discover when it was wrong, repair itself, preserve state, and resume after interruptions.
The important part is that I don’t want autonomy to mean “give a model a huge prompt and trust whatever it says.” The system is designed so that claims and actions leave evidence, important state exists outside the model’s context window, failures are preserved instead of hidden, and other agents can independently challenge or verify the work. Ideally, you should be able to reconstruct why the system did something and what evidence supported it.
In the screenshots, Codex is acting as the execution/proof agent, while two independent “spotter” agents audit the live run. They inspect checkpoints and evidence, look for unsupported assumptions, missing tests, inconsistencies, or incorrect conclusions, and send targeted questions back to Codex. Codex investigates those issues, falsifies or confirms them, makes repairs when required, verifies the new state, and creates another durable checkpoint.
So the loop is roughly plan → execute → test → challenge → falsify → repair → independently verify → persist → resume. As the architecture improves, more of that supervision is being performed by the agent system itself instead of by me manually.
The broader research goal is a general-purpose, high-assurance agent architecture, not just a better coding bot: something where models can be swapped or combined, specialized agents can divide responsibilities, and the surrounding system provides persistent memory, provenance, verification, recovery, and control. The screenshots are basically a live experiment in whether that kind of system can keep itself on track during a long-running real workload.