The line that stood out to me is the "start with no done" ledger state: a crashed run looks like a run that never started. The same gap shows up in agents, where the final message says something succeeded and nothing independent confirms it.
A travel-planning agent we tested replanned from Whitefield to Koramangala, found an inn at INR 4800, then tried hold_booking at INR 5900, a price no search had returned. The tool refused, but the flight hold still went through and the final answer said "Infeasible". Only the recorded trace showed a half-completed side effect.
So the ledger needs to come from what actually happened, not from the agent's summary. That's what we built AiEnsured agentest, an Agentic AI Testing platform, to judge from recorded traces. It has 165 tests:
pip install agentest-ai
Developers are welcome to enhance and contribute to it.