Three product shapes with three different promises. An assistant answers when asked. A copilot works alongside you inside the task, suggesting as you go. An agent goes away and does the job. Each sets different expectations about supervision, and choosing the wrong shape is a common product mistake.
Technical view
They differ in who initiates, how much context the system holds, and how reversible the actions are. Assistants require little integration and place all verification on the user. Copilots require deep integration with the working surface and depend on suggestion quality and interruption cost. Agents require tool access, permissions, budget control and audit, and shift the interaction from output review to trajectory oversight, which is a materially harder design problem.
Related
Part of AI Terminology Index, section: Product and Adoption.