Each tool covers one way agents fail in production: they forget, guess, drift, pass flaky tests or run untrusted code. All of them are open source, and the ones built on research link to the paper.
14 open-source tools7 backed by arXiv preprints10 papers in all
Five problems
The flagship five.
Memory, decisions, contracts, testing and security: the core of an agent you can check.
Design-by-contract for agents: preconditions, invariants, governance and recovery checked at runtime, so a violation is caught at the step that caused it.
Two instances of the same model failed together on 90% of missions where either failed — independence is not a safe assumption