M. Touheed: agentic AI breaks testing's three core assumptions

The essay's starting point is that most traditional enterprise systems were built around three assumptions: jobs finish quickly, retrying one is free, and the same input always produces the same output. Agentic workloads break all three, and Touheed writes that the difficulty is rarely the model — it's the surrounding infrastructure and management practices that assume properties these workloads no longer have.

A large part of his argument is about why companies don't see this coming. When a person chats with AI and checks each response, that person is the error handler, so the error was always there but stayed invisible. Automated agent workflows run headlessly, and a workflow that behaved reliably in interactive use can behave differently when a scheduler fires it 400 times overnight with nobody watching. Touheed's prescription is to identify every task the human was performing manually and specify which automated check or system takes over.

He frames cost the same way. Since cloud and model API usage is billed asynchronously on monthly cycles, compounded retry costs quietly accumulate and become visible only when the invoice arrives weeks later — which is why he wants spending tracked per completed unit rather than per call. On failures, he draws a parallel to data pipelines: agentic work is long-running, partially failing and expensive to re-execute, and teams coming from application development, where synchronous request-and-response is the norm, may discover that gap as a series of surprising incidents rather than as a skills need.