A new paper argues that aligning an AI agent with its user isn't something you finish before shipping. It's something you have to keep doing every time they talk.
Researchers published the paper, titled SPEAR, on October 7, 2026. It argues that most alignment work treats the problem as pre-deployment optimization: collect human feedback, learn preferences, fine-tune a model, ship it. SPEAR instead splits alignment into five ongoing pillars: Specification, how people express what they want; Process, how an agent decides whether to act, ask, defer, or pause; Evaluation, how people judge whether the agent actually succeeded; Adaptation, how the agent adjusts to a given user over repeated use; and Recalibration, how people adjust their trust and expectations in response.
Most alignment research still stops at fine-tuning, treating a model's values as fixed once it ships. Agents that manage calendars, write code, or act on a user's behalf operate in longer, messier interactions than a single preference snapshot can capture. Naming evaluation and recalibration as distinct design problems hands product teams a vocabulary for behavior that today mostly gets patched over with settings menus and support tickets.
Five pillars and a tidy acronym won't stop an agent from confidently doing the wrong thing, but at least someone is finally asking what happens after the demo.