QA and regression testing
An agent that checks a flow is only useful if its behaviour is constant. The moment it carries memory across sessions, a failing check has two possible causes and you cannot tell them apart. Teams end up re-verifying agent output by hand, which is the work the agent was supposed to remove.
Research and analysis
Prompt iteration on a live-web tool is guesswork. You rewrite the instructions, the answer changes, and two things moved at once — your prompt and the pages the model happened to read that minute. Without a baseline you are tuning on noise.
Content operations
Tone drift is invisible while it happens. Nobody edited the prompt; the assistant simply accumulated forty sessions of context and started writing slightly differently. By the time an editor notices, there is no baseline to point at and no way to say when it started.
Compliance and regulated review
A reviewer asks which instructions produced a decision eight months ago. On a mutable assistant that question has no answer: the prompt has been edited since, the thread shows the new one, and the sources the model read have moved on. That is not a documentation gap — it is a reason not to be allowed to use the tool.
Whichever one you are, the fix is the same.
Pin the agent to a version, run it in a fresh context, and diff the versions when you change something.