Give the AI Agent the Harder Job: Ambiguity Yields Better Results
Giving AI agents ambiguous, harder tasks—paired with rigorous verification—surfaces hidden bugs and tackles the backlog work engineers never had time for.
The author describes asking an AI agent to simply 'QA the redesign' of a website, only to watch it independently spin up a headless browser, simulate clicks on the live page, and uncover a real bug in the filtering feature—caused by a props-encoding mismatch between the WebAssembly module and its JavaScript bridge. Nobody asked for a browser; the ambiguous request led the agent to invent the method itself, illustrating that vague instructions given to a capable agent can yield both the result and an approach the user never thought to request.
The central argument is that people habitually scope their requests to what they already believe the tool can do, and thus never discover its actual ceiling. Since the cost of a failed attempt has dropped from an afternoon of work to a single sentence, the rational move is to routinely ask for things assumed to be impossible. This shift also unlocks a backlog of low-priority-but-valuable chores—like dead-code audits or generated 3D assets—that were never quite worth a human's time but now fit easily into an agent's workflow.
Crucially, the piece warns that creative exploration without verification is dangerous: the same agent that improvises a clever testing strategy can also falsely assure that something works when it doesn't. The recommended discipline is to pair ambitious, ambiguous requests with adversarial verification—driving the real code path and checking actual behavior. For engineers, the bottleneck has shifted from available hours to the willingness to ask for harder things and the rigor to confirm the results.