DooDooLamb News
Everyone is building AI agents. Who’s designing the stop button?
Brief published November 5, 2026 · Original source published November 3, 2026
Original reporting by Michał Krztoń at uxdesign.cc.
Automated brief. Verify important details at the original source.
What happened
Researchers at Thinking Machines Lab ran an identical prompt through a language model 1,000 times with temperature set to zero, a setting intended to produce deterministic outputs. The experiment revealed variability in token selection despite that constraint. The article frames this as evidence that AI agents capable of interpreting, deciding, and acting autonomously introduce unpredictability that conventional user flow design was never built to handle, raising questions about where and how designers should place meaningful intervention points.
Why it matters
Builders shipping agentic systems are inheriting a design gap: the tools and conventions developed for predictable user flows do not map cleanly onto systems that can act without step-by-step human confirmation. The Thinking Machines Lab finding suggests that even low-temperature configurations do not guarantee consistent behavior, which complicates assumptions baked into current agent architectures.
What to watch
Whether the design and engineering communities converge on shared standards for interruption and override mechanisms in agentic systems, and whether further research confirms the temperature-zero variability finding at scale.