In the quiet machinery of automated intelligence, OpenAI discovered that its Codex coding assistant had been consuming the resources of paying users not through any single dramatic failure, but through the accumulated weight of eight categories of systemic bugs — loops that never closed, agents that never stopped, and memory systems that never received the signal to rest. The company has patched the issues, reset affected quotas, and acknowledged what thousands of users had already sensed: that the invisible labor of AI agents is far harder to account for than a simple conversation. This momen