StudioFive ways an agent run can end: have you designed all of them?
The author argues agent designs focus too much on the happy path and proposes explicitly defining five run endings: success (verifiable done), purposeful giving up, budget exhausted, blocked needing a human, and failure. Each comes with concrete design advice, such as giving the model a legitimate way to refuse, setting limits ahead of time, and separating retryable from permanent failures. Runs should end in exactly one recorded state, and the share of each ending is useful signal.
The article breaks an agent run into five explicit end states with design tips for each, useful for auditing how your own system terminates runs.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.