Why AI Agents Fall Apart on Real Work

I am building Tandem, an open-source autonomous execution engine. Here is what I learned hardening it against the failure modes that make long-running agents unreliable.